This PDF file contains the front matter associated with SPIE Proceedings Volume 13272, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Aiming at the problem of low safety caused by the traditional ant colony algorithm without considering the terrain and obstacle locations when performing path planning, a safe path planning method for robots based on the improved ant colony algorithm is proposed. Ensure the safety of the path without increasing the path length, fully consider the terrain slope, orientation, distance from obstacles, and the direction of obstacles to design the warning pheromone, so that it interacts with the tracer pheromone to reduce the tracer pheromone on the dangerous path, and find the optimal path that integrates the path length and safety. The 2.5D map is built in Python language for simulation, and the cumulative elevation difference is reduced by 26.7% compared with the traditional ACO algorithm, and the cumulative distance between the path and the nearest obstacle is improved by 5.8%. the simulation results prove the effectiveness and superiority of the optimization algorithm.

To address the problems of high memory consumption, heavy computational burden, and difficulty deploying existing detection algorithms on edge devices, a lightweight approach for detecting abnormal driving behavior based on an updated YOLOv8 model has been presented. Firstly, the backbone network is replaced with the StarNet network, which has lower computational and memory access requirements. Second, utilizing an enhanced weighted bidirectional feature pyramid network with C2f blocks enhances the model's capability to capture multi-scale features swiftly. Finally, an asymmetric multi-level compression detection head is introduced to reduce the model size further. The experimental findings demonstrate that the suggested enhanced model achieves a 73.2% reduction in parameters, a 60.9% decrease in computational burden, and a 72.8% reduction in model size while retaining the same accuracy as the original YOLOv8 model. This significantly reduces hardware costs, demonstrating excellent lightweight effects.

Silkworm microparticle virus is a serious silkworm disease in sericulture. It has a direct threat to the production of silkworm cocoon and silkworm species. In the recognition of silkworm microparticle virus, traditional manual microscopy methods have problems such as high labor intensity, unreproducible results, and low recognition accuracy. This article introduces the YOLO deep learning algorithm to identify the silkworm microparticle virus, collect the original image with the microscope, and get the data set after data enhancement. In order to improve the training accuracy the CBAM attention mechanism is introduced based on the Yolov5s network structure. The training results on the silkworm microparticle virus picture data show that the average training accuracy mAp can reach 1.5% following the integration of the attention mechanism. Finally, for different microparticle virus pictures, the YOlOv5s-CBAM structure is better than the Yolo5vs in the identification of microparticles.

Due to the extensive intertwining of the Internet and information systems, cybersecurity risks have markedly increased, necessitating a shift from safeguarding information to comprehensive cybersecurity. In response to the challenges posed by the high expenses and inefficiencies of traditional feature selection techniques, this study introduces an optimized Honey Badger Algorithm with reduced computational overhead (EHBA), coupled with Random Forest (RF) technology, specifically tailored for feature selection in cybersecurity datasets. This method not only boosts the precision and effectiveness of feature selection but also diminishes reliance on data, bolsters the model's ability to generalize, and efficiently manages computational expenses. The model promptly captures the current cybersecurity landscape and forecasts security trends, thereby providing robust support in addressing intricate cybersecurity dilemmas. Empirical findings substantiate its substantial efficacy in mitigating the escalating cybersecurity concerns.

Steel is the cornerstone of human life and production, and its quality is crucial. Detecting defects on steel surfaces using AI has become increasingly important. Considering the complex structure and diverse defect types of steel surfaces, as well as issues like lighting and environmental interference, we propose a steel surface defect detection algorithm based on YOLOv8-DIC. The main improvements include: first, integrating the DCNv4 module, which introduces deformable convolution to effectively recognize geometric transformations of objects in spatial scales and sizes, thereby enhancing model accuracy and reducing model size; second, incorporating the IRMA attention module to better focus on features at different scales and improve detection accuracy by fusing multi-scale information; finally, integrating the CSPPC module to utilize partial convolutions for handling redundant feature map computations, reducing redundancy and memory access, significantly shrinking the model size while improving accuracy. Experimental results on the NET-DET dataset show an average precision mAP of 78.7%. The model size decreased by 16%, and the improved algorithm achieved increases of 4.8% in precision, 1.5% in recall, and 3.3% in average precision, with reductions in false positives and false negatives. This enhancement greatly enhances the capability of steel surface defect detection while meeting real-time requirements.

Traditional blood cell detection methods are limited by cumbersome procedures, subjective judgment, technical constraints, and difficulties in data processing and interpretation. These limitations lead to low detection efficiency, high error risk, and issues such as misdetection and missed detection. To address these problems, we propose a YOLOv8-EOC blood cell detection algorithm, optimized and improved based on the YOLOv8 framework. This algorithm designs a novel C2f_EMA module by combining the C2f and EMA mechanisms within the residual connections of the backbone network, thereby enhancing detection accuracy and feature extraction capabilities. Additionally, by replacing some convolution modules with OMNI-DIMENSIONAL DYNAMIC CONVOLUTION (ODConv), the algorithm improves feature extraction capabilities, particularly in overlapping blood cell regions. Finally, by incorporating the coordinate attention (CA) mechanism in the neck network, the algorithm enhances the recognition accuracy of small target platelets in the dataset. Extensive experimental results on the BCCD dataset demonstrate that compared to YOLOv8n, this algorithm improves the average precision of detecting three types of blood cells by 2.7%. Moreover, with only 3.06M parameters and a computational load of 7.9 GFLOPs, the detection speed and accuracy of this algorithm surpass traditional methods and other advanced algorithms. The experimental results indicate that this algorithm has high practicality for blood cell detection and meets the deployment requirements for low computational power while ensuring high accuracy

The swift and precise detection of vehicles plays a significant role in intelligent transportation systems. Current vehicle detection algorithms encounter challenges of high computational complexity, low detection rate, and limited feasibility on mobile devices. To address these issues, this paper proposes a lightweight vehicle detection algorithm based on YOLOv7-tiny (You Only Look Once version seven) called Ghost-YOLOv7. The width of model is scaled to 0.5 and the standard convolution of the backbone network is replaced with Ghost convolution to achieve a lighter network and improve the detection speed; then a self-designed Ghost bi-directional feature pyramid network (Ghost-BiFPN) is embedded into the neck network to enhance feature extraction capability of the algorithm and enriches semantic information; and a Ghost Decouoled Head (GDH) is employed for accurate prediction of vehicle location and species; finally, a coordinate attention mechanism is introduced into the output layer to suppress environmental interference. The WIoU loss function is employed to further enhance the detection accuracy. Ablation experiments results on the PASCAL VOC dataset demonstrate that Ghost-YOLOv7 outperforms the original YOLOv7-tiny model. It achieving a 29.8% reduction in computation, 37.3% reduction in the number of parameters, 35.1% reduction in model weights, 1.1% higher mean average precision (mAP), the detection speed is higher 27FPS compared with the original algorithm. Ghost-YOLOv7 was also compared on BIT-vehicle datasets as well, and the results show that this algorithm has the overall best performance.

In order to avoid tracking loss in complex scenarios with large-scale occlusion and multiple interfering factors, an object tracking algorithm based on combined discriminative correlation filter and re-detection was proposed. First, on the basis of discriminative correlation filter, a spatiotemporal regularization term was introduced and then the object scale estimation was improved, to improve the robustness of the tracker in complicated scenarios. Then, a reliability evaluation module consisting of two complementary methods was added to optimize the update strategy of the filter. Finally, a re-detection module was designed to execute two strategies for re-detection when the original tracking results are unreliable. After selecting the optimal re-detection results, the object tracking is completed. The experimental results show that, compared with the baseline tracker STRCF, the tracking precision of the proposed algorithm on OTB100 datasets and UAV123 datasets is improved by 0.9% and 4.3%, and the tracking success rate is improved by 1.5% and 5.1%. Real-time object tracking with fast running speed, good ac-curacy, high precision and strong robustness has been achieved.

To address the problems of low bubble testing accuracy and easy leakage of small bubbles in the gas-tightness inspection of stainless steel defective welded pipe by existing methods, an improved YOLOv8-CME algorithm based on YOLOv8 is proposed. The method replaces the C2f module with an improved C2f_CM module in the feature extraction and feature fusion network, extracts the image channel information and spatial information at the same time, and improves the detection accuracy while lightening the model structure. The ECA attention mechanism is introduced to improve the model's ability to detect fine bubbles and prevent leakage detection. Experiments are conducted on the self-constructed bubble dataset, and the data show that: compared with the original YOLOv8 model, the model FPS of the improved algorithm in this paper is improved by 12 f/s, the average precision is improved by 5.3%, the accuracy is improved by 3.7%, and the recall is improved by 3.4%. The experimental results verify the effectiveness of the YOLOv8-CME algorithm.

In view of the problems of tire rolling resistance and longitudinal wind resistance in vehicle driving, a planar four-wheel seven-degree-of-freedom vehicle dynamics model was established. For the purpose of realize vehicle platooning, a longitudinal following control method based on fuzzy PID was proposed, and the following control effect was analyzed on a sinusoidal path. The experimental results show that this method can enable the following vehicle to maintain a stable longitudinal desired driving distance, and has a certain improvement in trajectory following accuracy and formation stability.

Aiming at the problems of low accuracy and large number of parameters of the current general-purpose detection methods in coping with the tasks of complex background, part occlusion and multi-scale part recognition, a lightweight industrial part detection algorithm is proposed on the basis of YOLOv8s. Designing C2f-DRRCB efficient feature extraction module in the backbone network to improve the multi-scale feature fusion performance; by introducing a lightweight Adown downsampling module, more feature-critical information is retained while reducing the number of model parameters and computation; the EMA multi-scale attention mechanism is added at the end of the backbone network to improve the network's attention to the important feature regions and reduce the background information's interference; Design of a new lightweight inspection head, SCAD-Head, to improve accurate localization of part targets by capturing long range dependencies and shared convolution. The improved algorithm is applied to the SIP-17 industrial part synthetic dataset for validation, and the results show that compared with the baseline YOLOv8s model, the average detection accuracy of the improved model is improved by 5.7%, the number of parameters is reduced by 35.85%, and the amount of computation is reduced by 27.72%, which is also a certain degree of improvement in the detection performance compared with the other latest detection algorithms.

Smoking in public areas poses significant risks to both individual health and the well-being of bystanders. Given the microscopic nature of smoke particles, detecting them can be challenging. To address the problems of its low detection accuracy, easy to be obscured and ignored, we propose a segmentation approach based on YOLACT_SMOKE. In this study, we enhance the original YOLACT model by integrating the smoother Mish activation function, which optimizes model speed without sacrificing performance. Additionally, we replace the Res2Net module with the ResNet module to improve feature extraction and expression. Furthermore, we introduce the CBAM attention mechanism to the C3 and C4 layers to mitigate noise interference. Recognizing the lack of suitable datasets, we curated a self-made smoking dataset for training and evaluation purposes. Experimental results demonstrate a notable 4.8% increase in accuracy at IoU thresholds of 0.5 and 0.75 compared to the pre-upgrade model, with respective improvements of 5.3% and 6.7% over the original model.

In this paper, a novel active noise reduction algorithm SFxNLMS (sigmoid filtered-x normalized least mean square) is proposed, which aims to solve the problem of slow convergence of the traditional FxLMS (filtered-x least mean square) algorithm in the field of active noise reduction. Based on the FxLMS algorithm, the algorithm introduces a sigmoid function to nonlinearly process the filtered error signal, and adopts a normalization strategy to optimize the learning step size, thus improving the convergence speed and stability of the algorithm. Using Gaussian white noise as the input signal, the FxLMS, SFxLMS and SFxNLMS algorithms are used for the noise active noise reduction simulation comparison experiments to compare and analyze the performance of these three algorithms. The results show that compared with the other two algorithms, the average noise reduction and convergence speed after active noise reduction using the SFxNLMS algorithm are substantially improved.

In recent years, with the development of economic globalization in the post-epidemic era, the shipping market capacity has made steady progress, and the accompanying problem of black smoke pollution from ship exhaust is still one of the objects of concern outside the industry. In order to improve the efficiency and accuracy of ship exhaust black smoke target detection. In this paper, the backbone network of YOLOv5s is improved, and a ship exhaust detection model YOLOv5s-CAF is proposed. Firstly, a small target detection layer is added to retain more shallow semantic feature information. At the same time, a simplified weighted bi-directional feature pyramid network Light-BiFPN is proposed to replace the original path aggregation network PANet, which realizes multi-scheduling feature fusion. Finally, a CA (Coordinate Attention) mechanism is embedded in front of the detection head to enhance the network's ability to localize the ship exhaust gas area. The results show that the Precision and mAP0.5 of the improved algorithm YOLOv5s-CAF reach 91.4% and 83.9%, respectively, which improve 9.8% and 7.6%, respectively, compared with the original model, verifying the effectiveness of the improved strategy in this paper.

In response to the problem of low recognition accuracy for green pepper picking in intelligent picking scenarios, an enhanced model based on the YOLOv7 object detection algorithm was proposed to elevate the recognition accuracy of green peppers. Firstly, the Self Attention mechanism is introduced to focus on the correlation between different positions, thereby capturing richer spatial and channel information of contextual information and reducing interference from other targets. Secondly, ASFF feature fusion module is added to adaptively integrate multi-scale features by learning weights to improve target recognition ability. Through training experiments on the established dataset, according to the test results, the evaluation accuracy rate (mAP) of the improved YOLOV7 model is 92.7%, which is an increase of 3.2% compared with the original algorithm. This result proved that the improved model showed higher efficiency in identifying green pepper.

We propose a target detection algorithm that combines confidence and feature extraction to address the problems of existing target detection algorithms, such as low detection accuracy and susceptibility to occlusion. Based on the traditional algorithm, this algorithm introduces the adaptive strategy objective according to the motion complexity of the video block, combined with the target contour field retrieval to enhance the precision of the algorithm. Experimental results show that in different environments, the algorithm in this paper has better detection accuracy and anti-obscuration ability than other existing algorithms. And it still maintains a high resolution for multi-targets, which effectively reduces the problems of target loss, identity jumping, and so on.

Aiming at the problem that the original Canny algorithm is incomplete in detecting the edges of the image and is very sensitive to the pretzel noise and the small changes in the edge gradient, this paper proposes an edge detection algorithm based on the improved adaptive Canny algorithm. Firstly, a fast median filter is utilized in place of a Gaussian filter to accomplish the image denoising process. Next, the convolution template of the traditional Canny algorithm is extended to four directions to determine the magnitude of the gradient in order to prevent the loss of edge details. Finally, it uses maximum interclass variance (Otsu) to adaptively select the dual thresholds, which improves the accuracy. Experimental results demonstrate that the algorithm presented in this paper achieves superior edge integrity and noise reduction. with fewer pseudo edges, the denoising quality improves by 15-20% and the edge evaluation index improves by 20–25%.

DNA-Depth is currently one of the better-performing lightweight models for monocular depth estimation, but predicting detailed information remains challenging. This paper proposes a method called EUF-Depth, which can efficiently fuse encoder features based on DNA-Depth. Firstly, a feature fusion strategy is introduced to improve the utilization of important features through a learnable method. Then, the channel attention in the decoder part is replaced with Coordinate Attention (CA), incorporating positional information to enhance model prediction accuracy. Experiments on the KITTI benchmark demonstrate that EUF-Depth outperforms DNA-Depth in all metrics. The error evaluation metrics decrease by an average of 2% and the accuracy evaluation metrics increase by an average of 1%.

At present, the polar region is used to monitor and investigate fish by combining sonar detection with artificial fishing statistics, which is limited by economic cost, operation area and time. Object detection algorithms based on deep learning can identify and detect fish while meeting economic requirements, but traditional object detection algorithms often have many parameters and calculations and cannot adapt to the harsh conditions of energy consumption and storage limitations in the polar region. To solve this problem, an improved lightweight fish detection algorithm for YOLOv8n was proposed, in which the GhostC2f module was used to replace C2f in the backbone and neck networks, and GhostConv was used to replace part of the Conv in the network, and the EMA attention mechanism was introduced in the backbone network to improve the feature extraction ability. Finally, the MPDIOU loss function, which is simpler in the calculation process, was used to replace the CIOU to improve the detection speed. Experiments on the self-made fish dataset show that the number of parameters and computation of the improved algorithm become 1.49M and 4.7GFLOPs, respectively, and only 49.67% of the parameters of the original YOLOv8n are used to achieve the same detection accuracy, which meets the requirements of model deployment under limited hardware conditions.

General defect detection methods perform well in detecting obvious defects, but detecting subtle defects such as scratches and cracks in steel still poses a challenge. This paper proposes a steel defect detection algorithm based on the YOLOv7-tiny model[1], and names as CaGFPN-YOLO. Firstly, a CA module is introduced into the feature extraction network to enhance the model's ability to extract subtle defect features; Subsequently, the neck structure of the YOLOv7-tiny model is replaced with an improved RepGFPN structure to fully integrate multi-scale features, further enhancing the model's ability to detect subtle defects. Experiments conducted on the Northeastern University hot-rolled steel surface defect dataset show that the CaGFPN-YOLO model improves the mAP@0.5 by 3.1%, recall rate by 3.6%, and F1 score by 3.5% compared to the YOLOv7-tiny model, the CaGFPN-YOLO model can effectively enhance the detection capability of steel defects.

Addressing the challenges posed by varying shapes and sizes of maintenance hole covers in complex environments, as well as the high computational complexity and numerous parameters of existing models, this study presents an improved two-step YOLOv8 algorithm for detecting maintenance hole hazards. Firstly, the model structure was optimized by incorporating the Multi-Head Self-Attention Mechanism (MHSA), enhancing its capability to detect maintenance hole covers against complex backgrounds. Secondly, the Deformable Convolution Network version 2 (DCNv2) was applied to accommodate the diverse shapes of maintenance hole covers, further increasing the detection precision. Testing on a custom dataset showed that the improved two-step YOLOv8 model achieved a mAP50 of 0.984 at an intersection over union (IoU) threshold of 0.5, a recall of 0.95, and a precision of 0.964. Compared to the original YOLOv8 model, this model demonstrated significant performance improvements with increases of 0.052 in mAP50, 0.089 in recall, and 0.087 in precision. These enhancements sped up the detection of maintenance hole covers and improved accuracy, providing adequate support for practical applications. Future research will consider integrating this model with embedded devices to promote intelligent and efficient urban monitoring.

An enhanced DDE-DTS algorithm is proposed to solve the substation camera inspection work task scheduling problem. Firstly, considering the camera scheduling problem in actual inspection tasks, the maximization of the minimum completion time of inspection work is set as the objective, and constraints are established to develop an optimization scheduling model. Then, a search space is established using transformer sequences as arrangement codes, and an improved WPF-NEH joint rule is proposed to generate the initial population. Furthermore, in order to improve the efficiency of the solution, an enhanced discrete differential evolution algorithm is employed in the global parallel search phase, and a dynamic taboo search algorithm that takes into account both concentration and diversity is proposed for the local search phase, forming an enhanced hybrid DDE-DTS algorithm. Finally, the algorithm is successfully applied to optimize the scheduling of substation camera inspections in actual inspection work.

In recent years, video matting has become a hot topic across various industries, especially with the extensive use of portrait video matting. Therefore, achieving precise and rapid matting results has become a focal point of recent research. While conventional image matting can relatively accurately obtain precise foreground portraits, the results are still suboptimal for video matting. This paper uses video as the sole input, and by incorporating a triplet generation module and a spatiotemporal feature aggregation module, we embed a twin network into our architecture to achieve faster and more accurate portrait video matting under temporal consistency conditions. Comparative results demonstrate that, compared to the Deep Image Matting (DIM) method, our approach reduces MSE by 73.3%, SAD by 51.3%, Grad by 6.9%, Diff by 73.6%, and dtSSD by 87.6%.

There are issues with capturing the actions of moving objects in video sequences with sufficient accuracy. Introducing a three-dimensional convolutional neural network (3D-CNN) can effectively capture dynamic information in video sequences. However, due to its fixed receptive field, it cannot capture long-range dependencies. To address these issues, this paper proposes an algorithm that combines a three-dimensional convolutional neural network and a 3D self-attention mechanism. This method primarily improves action recognition accuracy by capturing spatiotemporal information in video sequences. Firstly, 3D-CNN is utilized to perform convolution operations in width, height, and time dimensions, extracting spatiotemporal features from video frames. Secondly, the self-attention mechanism is extended to three dimensions by introducing a 3D self-attention mechanism to process spatiotemporal data. This mechanism calculates the relationship between each frame in the video sequence and all other frames, capturing long-range dependencies. Experimental results show that, compared to other algorithms, the recognition accuracy of the proposed algorithm increases by at least 7.2% and 5.4% on the UCF-101 and HMDB51 datasets, respectively. Additionally, the proposed algorithm has good interpretability, allowing for understanding of the model's decision-making process through visualizing attention weights.

This paper proposes a wind turbine fault prediction model based on genetic algorithm optimization. Firstly, a large amount of wind turbine operation data is collected and feature engineering processing is completed. Secondly, a fault prediction model is established as a benchmark model based on BP neural network. Again, using genetic algorithm to optimize the parameters of this model to improve its prediction performance can effectively improve its prediction performance. The genetic algorithm continuously optimizes the model parameters to find the optimal solution by simulating the natural biological evolution process. Finally, the optimized model is experimentally verified and compared and analyzed with the traditional method. The experimental results show that the wind turbine fault prediction model optimized based on genetic algorithm achieves significant improvement in both accuracy and generalization ability. Compared with the traditional method, the optimized model is able to predict wind turbine faults more accurately, discover potential problems in advance and take corresponding maintenance measures, so as to improve the reliability and efficiency of the generator.

In the field of computer vision, image classification is a very important branch, it‘s aims to classify different images into different category labels. Traditional image processing methods improve generalization ability and model accuracy by learning large amounts of data, but in practical applications, acquiring large amounts of labeled data is often very difficult and costly. Therefore, small sample Learning and meta-learning have become effective methods to solve this problem. This paper discusses its application in small sample image classification from the perspective of auxiliary meta-learning.

The crystal size is an important parameter for monitoring and control in industrial crystallization processes. In this paper, we put forward an optimized measuring technology for size measurement by using computer vision methodologies. The focus is to overcome the uneven background with filtering and deblurring methods, image segmentation, and size measurement. The crystals can be segmented from the crystallization images by the improved combination method. Then the particle sizes are obtained by fitting the crystal shapes for analysing the particle size distribution. The experimental results present the proposed model has high effectiveness based on real-time crystallization images.

To address issues such as incomplete detail restoration and partial loss of structural information during the dehazing process, a single image dehazing algorithm based on multi-scale feature fusion and interaction was proposed. The algorithm based on U-Net network architecture and utilizes heterogeneous convolutions to extract and fuse multi-scale features, thereby effectively reducing the loss of image detail features during down-sampling. Simultaneously, during image reconstruction, the paper proposes an image enhancement module constructed based on feature fusion and interaction, which improve the reconstruction accuracy and generalization of the model. Experimental results show that the proposed network achieves an SSIM of 0.9896 and a PSNR of 35.89dB on the SOTS-outdoor dataset. Compared to dehazing algorithms such as MSBDN, De-Hamer, and gUNet-s, the proposed network's SSIM and PSNR have average improvements of 0.04 and 6.02dB, respectively.

Breast cancer is one of the common cancers worldwide, and how to identify breast cancer as early as possible plays an important role in its later diagnosis and treatment.However, relying only on data from a single organisation can limit the ability of trained neural networks to meet the performance demands of medical diagnostic services, and centralised learning that collects data from decentralised organisations can lead to privacy leakage. To address the above challenges, this paper proposes an assisted diagnosis model for breast cancer combining federated learning and improved DenseNet. Specifically, in contrast to mainstream centralised learning, this paper introduces federated learning to integrate multiparty data and effectively prevent privacy leakage. Each client sends the local model to the central server for aggregation, and then the global model obtained from the aggregation is sent to each client for training,thus achieving collaborative training of neural network models.In addition, in order to enhance the ability to capture feature information, this paper incorporates a Dynamic Convolution module into the densely connected network to improve the diagnostic accuracy without significantly increasing the size of the network model. The experimental results show that the proposed method has better diagnostic performance in breast cancer diagnosis compared with other methods.

With the rapid development of artificial intelligence technology, computer visual question and answer technology as well as scene description analysis technology are becoming more and more mature, and have been widely used in various fields of society. However, the current visual navigation scene analysis often suffers from problems such as incomplete scene description and inaccurate visual question and answer. A method combining the attention mechanism and visual Q&A is proposed, and the method is applied to navigation scene analysis to construct a visual navigation scene analysis model based on visual Q&A and attention mechanism. The model is subjected to performance comparison experiments using the similarity between the generated answer and the reference answer and the accuracy rate of scene description as the indexes, and the experimental results show that the average similarity and the average accuracy rate of the model are 0.598 and 87.6%, respectively. The above results show that the performance of the model is significantly better than the comparison model and is effective.

Aiming at the problem that the existing image quality evaluation indexes are sometimes inconsistent with the visual perception of the human eye, this paper proposes an image quality evaluation method based on a reference image-guided multiple path fusion network. Firstly, the fused image (the image to be evaluated) and the reference image are inputted into two network branches respectively, and each network branch contains a Transformer module (VIT-Block), a Convolution module (CNN-Block) and a Center Difference Convolution module (CDC-Block), which are used for extracting the remote information, retaining the background information, and extracting the texture features, and the reference image is used as the output of each module of the fused image branch, and the reference image is used as the output of each module of the fused image branch. The output of each module of the branch is used as the weight of the output feature maps of each module of the fused image branch, and the weighting process is carried out to obtain three groups of feature maps, and then the dimensionality reduction process is carried out by using two 1×1 convolutional layers respectively to obtain three significant feature maps, and then the global average pooling is carried out respectively to obtain three evaluation scores, and finally, the average of the three scores is taken as the final objective evaluation index value of the fused image. The experimental results show that the results obtained by the image evaluation method proposed in this paper are consistent with the subjective visual perception.

Light field data provides effective spatial information and focus cues for saliency detection in complex scenes. However, existing research on light field saliency detection still faces limitations. Therefore, this paper proposes a light field image saliency detection network based on Multi-Modal Feature Cooperation Fusion. First, we construct a Focal Feature Aggregation (FFA) module, which aims to explore and enhance features obtained from the unique structure of focus stacks. At the same time, the Multi-Modal Feature Cooperation Fusion (MFCF) module is utilized to establish connections between features in two different dimensions, achieving multi-level complementary effects through cross-modal feature collaboration between all-focus and focus stack features.

Continuous acquisition of the latest information about the shape of the object allows for more efficient and robust classification, as well as accurate estimation of the target state. However, previous methods have often overlooked this problem and used only the target information from the first frame in tracking. In this paper, we propose three specific and practical guidelines aimed at updating the target state, enabling the development of an anchor-free generic object tracker without requiring any prior knowledge. These guidelines offer a clear path and direction for the development process. Using these guidelines, we develop our Dynamic Update Template (DUT) tracker that includes a template, a dynamic template, and a search branch, ensures unambiguous classification scores, provides estimation quality scores, and multiplies them to obtain the pcore, which serves as the basis for updating the dynamic template. By conducting thorough analyses and ablation studies, we validate the efficacy of our proposed guidelines. Our DUT tracker achieves better performance on challenging benchmarks (LaSOT) without excessive modifications. On the extensive TrackingNet dataset, DUT attains an impressive AUC score of 82.1 while maintaining a swift frame rate exceeding 90 FPS, surpassing the threshold for real-time performance.

With the deep integration and wide application of Artificial Intelligence (AI) technology in medical imaging, medical imaging is experiencing unprecedented technological innovation and clinical practice breakthroughs. This study focuses on the efficacy of continual learning (CL) under the federated learning (FL) framework in the medical image recognition task, aiming to simulate the collaboration between different organizations across regions in a real-world healthcare environment. Considering data privacy is paramount in healthcare systems. This study proposes an FL scheme that upholds data privacy regulations while enabling CL with dynamically added medical image datasets. This Experiment utilized diverse medical image datasets across five human body regions, encompassing gastrointestinal diseases, brain lesions, and ophthalmic disorders and so on. Through in-depth experiments, we compare and analyze the performance of three federated class continual learning (FCCL) strategies, Finetune, FedEWC, and FedLwF. The analysis focused on their ability to handle emerging classes while maintaining recognition accuracy for existing ones, and explored the underlying reasons for performance differences. Comparative analysis revealed that FedEWC and FedLwF exhibit different CL capabilities by virtue of the regularization technique and the knowledge distillation transfer learning mechanism, respectively. Among them, FedEWC shows excellent performance stability in stomach and brain image recognition, while FedLwF demonstrates better performance than FedEWC on eye diseases and full-volume datasets through knowledge distillation.

The significant progress of vision-language pre-trained models (VLMs) and large language models (LLMs) have provided a feasible new mode for image captioning, which relies on VLMs to process images and then utilizes LLMs to generate captions, simplifying the caption generation process and making it lighter. Based on this mode, to address the issues of deviation between the generated captions and the content of image expressions, incomplete descriptive information, and resource hungry in previous Chinese image captioning models, we propose a Chinese image captioning model via double decoding based on Visual Prompts, DDVP. This model employs the CLIP model as the encoder, the GPT2 model as the decoder, and introduces Visual Prompts, which are keywords related to image content. The model adopts a double decoding approach, first decoding to generate Visual Prompts and then decoding to generate the final captions based on the Visual Prompts. Through evaluation, we have demonstrated that our model achieves competitive results on the AIC-ICC dataset, and while maintaining fluency, the generated captions of DDVP can also cover the information in the image more comprehensively and accurately.

This paper, originally published on 03-10-2024, was withdrawn on 21-10-2024 per author request.

Current deep learning-based omnidirectional image quality assessment (OIQA) methods essentially rely on the global features of omnidirectional images and pay little attention to the local features, and most of them have high computational costs. To solve these problems, a light-weighted model for omnidirectional image quality assessment is presented to accomplish the accurate assessment by extracting and fusing features in spatial and gradient domain. Therefore, spatial attention features are obtained from the spatial information of viewport images as global features, while gradient features are obtained from gradient information as local features. These features are then combined using a Multi-Modal Feature Fusion Network to improve the ability of the model in regards to the feature representation. Experiments were conducted on the two datasets, namely, the CVIQ and OIQA datasets to evaluate the proposed method. The experimental outcomes indicate that the proposed method has high competitiveness in the aspects of performance and model complexity compared with other representative methods.

The Knowledge Graph (KG) is a model that represents structured knowledge by capturing entities and their relationships in the real world. It is widely used in search engines, recommendation systems, and natural language processing. Knowledge Representation Learning (KRL) transforms semantic information from knowledge graphs into continuous vector space representations, thereby improving knowledge acquisition and reasoning capabilities. However, current KRL faces issues such as underutilization of entity attribute information and difficulty in handling zero-shot scenarios. This paper proposes the Integrated Embedding Model (IEM) to address these issues. IEM employs a BERT-based attribute encoder and attention mechanism to weigh different attribute types, creating reliable attribute information embedding. During training, IEM merges structure and attribute representations, demonstrating excellent performance in knowledge graph completion tasks. Additionally, this paper introduces the Open Domain Representation Learning Model (ODRLM) for entity and link prediction in open domain knowledge graphs. ODRLM enhances the representation of zero-shot entities and relationships through three stages of optimization. Experimental results show that this model significantly improves both entity and relationship prediction accuracy, effectively addressing challenges in knowledge graph completion, especially in Zero-shot scenarios.

Automatic classification of pork primal cuts holds significant application value in the food processing industry. To improve the accuracy of pork primal cut image classification, this paper proposes a classification model based on an enhanced EfficientNet-B0 architecture. Firstly, a pork primal cut dataset is constructed, covering various common cuts such as pork belly and pork loin. Secondly, to improve the performance of existing models, a Dynamic Channel Attention Module (DynaCAM) is designed and integrated into EfficientNet-B0. Experiments are then conducted on the self-constructed pork primal cut dataset and miniImageNet. The results demonstrate that the enhanced model achieves optimal classification performance on the pork primal cut dataset while maintaining low parameter count and computational complexity. Although its performance on miniImageNet is inferior to certain large-scale models, considering both accuracy and efficiency, the proposed method still possesses significant application potential and provides new insights for the intelligent upgrade of the food processing industry.

The goal of the low-light picture enhancement challenge is to increase an image's visibility without sacrificing its visually authentic appearance. This study addresses the issues of uneven brightness and negligible features in low-light photographs by proposing an enhanced Retinex algorithm for low-light image improvement. The RGB image is first converted to an HSV color space, and the multi-scale Retinex (MSR) algorithm is then applied to the V channel (brightness channel). This algorithm replaces the Gaussian filter in the Retinex algorithm with an adaptive bilateral filter and a Gabor filter, combining the qualities of the two filters to improve the image's brightness and details; Then, the original image is sharpened to improve the details of the image; Finally, the enhanced image is weighted and fused with the sharpened image to avoid the loss of details in the dark or overexposed regions of the image. The methods described in this paper are significantly improved compared with the classic Multi-Scale Retinex (MSR) algorithm and Multi-Scale Retinex with Color Restore (MSRCR) algorithm, which are able to repair the appearance of the image while maintaining its color and edges. color and edges while restoring the look and feel of the image.

Aiming at the existing multi-exposure image fusion method based on weight function, the fused image obtained tends to introduce halo artifacts, loss of details, unclear edges and other problems. A multi-exposure image fusion method with improved weight function is proposed. First, the luminance weights reflecting the importance of each pixel of the image to the global luminance and neighboring exposure images are determined, and the gradient weights reflecting the importance of the local gradient of each pixel in the whole sequence of the image are determined, and the contrast weights are used to improve the clarity and detail information of the integration results; finally, the fused images are obtained according to the weights. According to the experimental results that the image fusion effect of the method proposed in this paper is clear and natural, and the average multi-exposure fusion structural similarity (MEF-SSIM) reaches 0.981, and the average running time is 0.725s.

Aiming at the problem of arc image with blurred edges and inhomogeneous greyscale, a multi-feature-based active contour model for arc image segmentation is proposed in this paper. Firstly, based on the greyscale feature, the global and local based active contour model is established using the intra-class variance minimization criterion; then, based on the gradient feature, the second-order LoG operator is introduced and the optimized LoG energy is combined in the energy functional. The experimental results show that the proposed model outperforms some existing representative segmentation methods in terms of segmentation quality and efficiency, which verifies the effectiveness and robustness of the proposed model, and can provide technical support for the vacuum arc research.

Binocular stereo vision can perceive the stereo information of three-dimensional scenes under various conditions. Distance measurement based on binocular stereo vision plays an increasingly important role in distance measurement due to its unique advantages. The key technologies such as the principle of binocular distance measurement, camera calibration, stereo correction, and stereo matching are studied. The binocular vision information system is used to realize the ranging of stationary objects in 3D space. The main construction of the experimental model and the SGBM algorithm are used to realize the image matching operation when the object generates the disparity map, and the experimental results prove that the proposed algorithm can effectively measure the object size and the distance between the object and the camera optical center, and accurately measure the 3D position of the object at the same time. In addition, point cloud information can be generated for reference image 3D reconstruction.

The proposed super resolution reconstruction method aims to address the issues of missing detailed information in low light environments and poor recovery of high frequency details in reconstructed images. To fully extract the retained features of low-light images, a multi-scale feature fusion learning network is introduced, which utilizes convolution to extract shallow features and then extracts deep features through multiple cascaded multi-scale feature extraction modules. Additionally, an attention mechanism is incorporated to adaptively learn the importance of each channel and finally fuse the obtained shallow and multi-scale features. Furthermore, perceptual loss and edge loss are integrated into the loss function to optimize the network for high frequency texture details in the recovered image. Experimental results demonstrate that this method yields clearer edge structures and texture details after super-resolution reconstruction of images acquired in low light environments, outperforming other existing methods both visually and objectively.

In the image of minor defects in the appearance of brake pads, there may be significant differences in the number of different types of defects, resulting in a small sample size for some categories, making it difficult for the model to learn the characteristics of these categories of defects. To this end, a small sample learning based image detection technology for brake pad appearance defects is proposed. By using the Canny operator to obtain local maximum gradient amplitude pixels in the image of minor defects on the appearance of brake pads, edge detection of the defect image is achieved, and defect shape features are extracted. Using Boosting algorithm to construct a small sample learning model, calculating different weights for different training samples, to adjust the distribution of training samples for subsequent weak learners, in order to reduce overfitting and underfitting problems during the small sample learning process, and achieve image detection of small defects in the appearance of brake pads. The experimental results show that the method proposed in this paper has a high accuracy and recall rate for detecting small defects in the appearance of brake pads, including edge graying, uneven wear, cracking, and edge wear. Moreover, accurate detection of small defects can be achieved under small sample conditions.

The continuous advancement of virtual reality and Internet technology has gradually integrated 360° video into daily life. With its characteristics of high immersion and interactivity, 360° video offers an immersive viewing experience that has garnered widespread attention. However, the rich content inherent in 360° videos results in large amounts of data, imposing significant demands on the network environment required for video transmission. Given the fixed field of view of headmounted displays (HMDs), theoretically, only a small part of the video content needs to be guaranteed to ensure a seamless viewing experience and significantly reduce bandwidth consumption. Therefore, accurate viewport prediction is crucial. In this paper, we propose a 360° video viewport prediction framework based on multi-feature fusion using Transformers to address the limitations of existing methods, which often focus on a single type of data and exhibit low long-term prediction accuracy. Our approach leverages three different types of data: the user's past head movement trajectory, video saliency, and cross-user interest areas, to achieve more robust viewport prediction. Experimental results demonstrate that, compared to existing methods, our approach more accurately predicts future viewport areas and significantly improves the accuracy of long-term predictions.

The segmentation of retinal vessels is of paramount importance for the diagnosis of diseases and the performance of surgical procedures. With the advent de novo of in silico technologies, the field of in vivo vascular segmentation has witnessed a veritable explosion of innovation. In traditional segmentation networks, the challenge lies in the unclear delineation of vessels, which frequently results in a loss of important data. In response to this quandary, we present an innovative technique for vessel segmentation as part of our investigative work—the Lightweight Fully Connected MultiScale Attention Network (SCB-UNet). The use of data augmentation allows for the acquisition of a substantial quantity of annotated data without the necessity of utilizing thousands of datasets. This network combines the Full-Resolution UNet (FR-UNET) and the Full-Scale Cooperative Network (UNET3+) while integrating spatial and channel attention modules (CBAM). The network expands through high-resolution convolution interactions while preserving image resolution. Within the network, neighboring features are fused, and features from previous images are incorporated to gather contextual information. To enhance feature maps, attention fusion modules are introduced, in order to enhance the expressiveness of feature maps, it is necessary to concurrently strengthen them across spatial as well as channel perspectives. The efficacy in vivo of a novel network was evaluated on retinal vessel datasets (DRIVE, CHAIN_DB_1, & STARE). The results demonstrate in vivo efficacy superior to that of the aforementioned methods, thereby showcasing the network’s status as a highly efficacious and state-of-the-art approach.

Equipment demand prediction is the basic work of equipment guarantee, and it has traction for equipment financing and replenishment. Using the projection tracing regression model, the prediction model of the factors affecting equipment consumption, such as the operation type style, operation duration and task difficulty, the comprehensive index of detachment behavior, the intensity of the operation, the number of personnel, the operation equipment, and the field conditions, etc., has been set up, and the experiments have shown that the prediction accuracy of this model is high, which solves the problem of the previous low prediction accuracy by using only historical data and traditional experience. The model has high prediction accuracy and solves the problem of low prediction accuracy by using only historical data and traditional experience, which is of great significance to the equipment guarantee.

The accuracy of traditional visual flow measurement is often compromised due to the variation in the particle state of the material being transported by the belt conveyor. To address this issue, this study proposes a material flow measurement algorithm based on orthogonal machine vision. The algorithm begins by identifying the material type in the RR-CNN image, followed by establishing the material equivalent structure and visual measurement structure. Parameters are then calculated based on the movement state of the material. To improve the filtering effect, K-means clustering filtering is employed, along with a seed region growing algorithm that enhances particle shadow characteristics and image edges. Additionally, volume measurement of different filling densities is achieved using a differential grid and three-dimensional micro-elements, enabling the calculation of the mass of the collected material. Experimental results demonstrate clear particle characteristics and edges, resulting in cost savings for material measurement and a visual measurement accuracy exceeding 95%.

As ocean exploration deepens, efficient and accurate identification of underwater targets becomes particularly crucial. However, traditional methods and current technologies face challenges such as scarce samples and complex imaging conditions when processing side-scan sonar images. Given the current state of limited sample augmentation methods and low image resolution for side-scan sonar, this paper improves upon the SRCNN method by integrating the CBAM attention mechanism and Perceptual loss function. This approach mitigates the issue of increased noise typically associated with conventional image super-resolution reconstruction, thereby enhancing the accuracy of the side-scan sonar target detection model. Consequently, this method has been proven to enhance the quality of super-resolution reconstruction of side-scan sonar target images, offering a new approach to improving the construction of underwater target detection models.

In terms of computer vision, image classification plays a crucial role. With the progress in deep learning technology, there has been a significant improvement in image classification performance. Due to their capability of self-learning, deep learning models are applicable to automatically extract features from a vast amount of data and effectively address the complexity in image classification by navigating the nonlinear relationship inherent in high-dimensional data. In this paper, a novel approach to image classification is presented on the basis of an enhanced VGG network, with the GELU activation function incorporated into the VGG network to address gradient vanishing for deep learning. Furthermore, the Efficient Multi-scale Attention Module (EMA) is introduced for the network to focus more on the retention of information in each channel. This is achieved by dividing the channels into subgroups and balancing the spatial semantic features within the feature groups. Through cross-dimensional interaction, this network can be applied to encode the global information and enhance the features. Finally, an experiment is conducted on the CIFAR-100 classification dataset to demonstrate the superiority of the proposed method in image classification.

Human action detection in videos is crucial across various fields, yet manually annotating video data is time-consuming and results in limited datasets, weakening model generalization. To address this, this paper introduces a semi-supervised dual-channel spatio-temporal contrastive learning method-STCL, for accurate human action detection in videos. By constructing a dual-channel model to extract and compare spatio-temporal features in videos and adopting a semisupervised learning strategy that enhances unlabeled samples through forward and reverse playback, using a contrastive model to compare features, the method learns model parameters to distinguish between different video samples. This approach leverages the spatio-temporal information of unlabeled videos, avoiding extensive manual annotation. Experimental results on the Something-V2 dataset show that, even with only 5% labeled data, our method significantly outperforms existing techniques. This demonstrates the effectiveness of the proposed STCL model in utilizing video spatio-temporal information, achieving good results even with limited labeled data.

Scene Graph Generation (SGG) aims to bridge images and text. Addressing the issues of data distribution imbalance and long-tail effects in existing SGG methods, we propose a new Multi-feature Fusion Embedding Network(MFE-Net). This network optimizes relationship prediction through multi-level visual feature fusion and hard-constrained semantic feature classification, aiming to reduce inter-class confusion and strengthen intra-class consistency. Experiments on the Visual Genome dataset demonstrate that our model significantly enhances the accuracy of relationship recognition, achieving a new level of performance. The research not only optimizes SGG performance but also provides new strategies for addressing classification problems involving easily confused relationships.

To obtain fast and accurate crop distribution information, the feasibility of using active remote sensing imagery (Sentinel-1A) and passive remote sensing image(Sentinel-2) to extract the winter wheat spatial distribution was analyzed. Firstly, Sentinel-1A winter wheat image at the whole growth stage were synthesized according to the phenology of winter wheat, and high quality Sentinel-2 images of winter wheat after overwintering were synthesized according to the normalized vegetation index(NDVI) time series curves of various types of features. Three classification schemes, Sentinel-1A images, Sentinl-2 images and fused Sentinel-1A and Sentinl-2 active-passive remote sensing images, were designed, and then winter wheat was classified according to the random forest algorithm on the Google Earth Engine (GEE) cloud platform. The results showed that, accuracy of the user and winter wheat producer accuracy based on Sentinel-1A images at the whole growth stage were 83.15% and 86.44% respectively, and there was more "pepper" noise in the extraction results; the user accuracy and winter wheat producer accuracy based onSentinl-2 images after overwintering was 87.98% and 84.75% respectively, and improved extraction accuracy compared with that of Sentinel1A image at the whole growth stage, but the classification results were influenced by the "same spectrum of foreign matter", resulting in many misclassifications; the user accuracy and winter wheat producer accuracy with fused active and passive remote sensing images were 96.57% and 95.48%, respectively, compared with that of using only a single source of data, the accuracy of classification of winter wheat was improved to different degrees.

Digital image processing, also known as computer image processing, refers to the process of converting image signals into digital signals and processing them using computers. Digital image processing first appeared in the 1950s, when electronic computers have been developed to a certain level, people began to use computers to process graphics and image information. Digital image processing as a discipline was formed around the early 1960s. The purpose of early image processing is to improve the quality of the image, which is aimed at the human object and to improve the human visual effect. In image processing, the input is a low quality image, and the output is an image with improved quality. Commonly used image processing methods include image enhancement and restoration, coding, compression, and so on. This paper is based on the new big data background algorithms based on the support of the optimisation of the image processing technology of computer programs. The future application of digital image processing is becoming more and more widespread, the status is becoming more and more important, in the future modernisation will play an unparalleled important role.

In order to create more integrated and realistic indoor space design, we propose the idea of constructing an indoor space design system under the application of virtual reality technology. By combining interactive development technology standards on online platforms, we have successfully achieved visual design of indoor space design systems. We have completed the process of interior space design using techniques such as spatial model construction, original room type modeling, model import, and embedded integrated mapping. The main content of this system includes using a model library and user needs to select different models and generate initial scenarios; Users can roam in this system, and we modify the plan based on their feedback and generate the final plan. This system provides a convenient platform for designers and users to observe design results. By combining 3D modeling technology and virtual reality technology, we have achieved excellent indoor space design results, enriched and expanded the content of indoor space and landscape design, and also provided users with a more convenient experience. This has played a crucial role in promoting the development of interior space design.

Fiber orientation distribution (FOD) is crucial for resolving complex fiber structures. Existing methods fail to fully utilize global context when reconstructing FOD, leading to inductive bias. Furthermore, they typically ignore the spatial positional relationships of FOD, resulting in limited global continuity captured. This study addresses these issues by introducing a novel network called FOD-Mixer. The FOD-Mixer uses the capability of the multi-layer perceptron to extract long-range dependencies, thereby circumventing inductive biases. In addition, we introduce a dynamic-mixer attention mechanism, which simulates more intricate spatial relationships by dynamically computing the mixing matrix. Furthermore, we propose an attention-based pooling module to filter out noise information. Experimental results demonstrate that the global contextual and spatial information provided by the FOD-Mixer can significantly enhance reconstruction tasks. The attention-based pooling module also effectively filters noise, further augmenting the reconstruction accuracy of the FOD-Mixer. Our research substantiates the advantages and feasibility of the FOD-Mixer in reconstructing high-quality FOD, offering a reliable reference for clinical applications.

Diffusion magnetic resonance imaging (dMRI) is a noninvasive technique used to quantify brain microstructure and reconstruct white matter pathways. In particular, diffusion-weighted imaging (DWI) is a widely available method to captures water molecule diffusion to generate images reflecting tissue microstructure differences, essential for clinical diagnosis and research. However, DWI faces challenges like reduced signal-to-noise ratio (SNR) due to signal attenuation. Enhanced SNR and denoising performance can improve the accuracy of diagnosing and monitoring neurological disorders, such as Alzheimer's disease and multiple sclerosis, by providing clearer and more reliable imaging data. Due to the high cost of DWI data acquisition makes obtaining clean images for ground-truth infeasible through repeated averaging Therefore, supervised learning frameworks based on clean-noisy image pairs are not applicable. To deal with these issues, we propose an innovative Cluster-based Semi-Supervised Learning Framework and Blind Spot Dense Fusion Network (CSSLF-BSDFNet). CSSLF utilizes the redundant characteristics of DWI data to alleviate the data acquisition problem for the training of deep learning models towards denoising tasks without the need of noise-free ground-truth. BSDFNet integrates concepts from blind spot networks, dense networks, and fusion networks. This approach effectively extracts spatially correlated noise features from DWI data, efficiently removes noise, and preserves the organizational structure. Experiments on the DWI dataset of the Human Connectome Project (HCP) demonstrate that our proposed framework and network significantly improve the SNR of DWI images, outperform existing state-of-the-art deep learning denoising models, and are capable of effectively reducing image noise while preserving the original tissue microstructure.

Fiber optic composite overhead ground wire (OPGW) bears the important task of transmitting power data and guaranteeing the safety of the power grid, but OPGW cables in the process of operation will often be due to bird pecking, wind dance, ice and other anomalies that affect the normal operation of the cable. Therefore, the identification of abnormal events affecting the operation of fiber optic cables is crucial to the operation of the power system. However, due to geographic and environmental factors, the collected cable anomaly data is often unbalanced, which affects the performance of the classifier and leads to misclassification. Aiming at such problems, this paper proposes a data optimization method that combines improved K-means and Borderline-SMOTE. After calculating the data density and Euclidean distance of the fiber optic cable vibration data collected by phase-sensitive optical time-domain reflectometer (φ-OTDR), the data is divided into clusters by using the improved K-means algorithm, and the samples of a few classes that are easy to be misclassified at the decision boundary of each cluster are identified, and the samples of such classes are classified by using the improved K-means algorithm. A few classes of samples are identified at the decision boundary in each cluster, and such samples are oversampled using the Borderline-SMOTE method to balance the data volume of various types of anomalous events. The experimental results show that the accuracy, recall and F1 score of the three classes of machine learning algorithms combined with this method are improved by more than 5% compared with the traditional methods.

Bridge deflection is a crucial parameter in diagnosing bridge structural health, and accurately predicting deflection is vital for ensuring bridges operate safely and effectively. Traditional deflection timing prediction methods may be affected by future information. To address this issue, this paper proposes an improved timing prediction Transformer (TSTF) model, which introduces causal time multi-head self-attention mechanism (CT-MSA) to ensure that the model only relies on past information when forecasting. The multi-scale time dependence is extracted by multi-layer window size setting method. This paper also introduces KAN (Key-Attention-Neural) full connection layer in TSFT model to further improve the accuracy and stability of deflection prediction. Data sensitivity analysis is also performed to test the influence of input data at different locations on model predictions. In the experiment, the long-term historical deflection data of the bridge is preprocessed. By comparing TSFT model with traditional time series forecasting methods, the TSFT model based on CT-MSA and KAN fully connected layer has better performance in forecasting accuracy and computing efficiency.

The quality characteristics and deviations of an aircraft are extremely important for the design and manufacturing of the aircrafts, which are important inputs for the aircraft propulsion and control systems. Therefore, before participating in flight tests, the aircraft must undergo mass characteristic measurements to obtain of the quality characteristic parameters of the mass, the center coordinates of the mass, rotary inertia and product of inertia for aircrafts. This article proposes a method for evaluating the quality characteristics of aircrafts based on Monte Carlo random error generation. Large scale simulation data of the quality parameters of various components of the aircrafts are obtained based on the random error generation. The relationships between the quality deviation of each component of the aircraft and the overall quality deviation of the aircraft is established through data mining between the characteristic parameters of each component and the overall characteristic parameters. The quality parameter evaluation of each component with errors is obtained through Monte Carlo simulation, achieving high-precision evaluation of the overall rotational inertia and inertia product of the aircraft.

Parkinson's disease, prevalent among middle-aged and elderly individuals[1], is a degenerative neurological condition characterized by symptoms such as bradykinesia, tremors in the extremities, and overall body rigidity[2]. As the population ages, there will be more and more cases[3].This disease develops from a complex interaction of genetics and environment[4].The accurate quantification of tremors is crucial for the management of this disease. However, tremor amplitude can be influenced by various factors, including environmental and psychological impacts, leading to erratic fluctuations. In clinical settings, physicians are tasked with rapidly evaluating tremor severity, often relying on their clinical expertise for manual labeling. The ability to effectively quantify the severity of Parkinson's tremors is vital for elderly patients and their caregivers. This paper addresses the existing limitations and employs suitable strategies to investigate the automated quantitative assessment of Parkinson's tremor.

In social recommendation, user's interest preference is influenced by multiple factors. It is a challenging problem to mine data information effectively in multiple types of interaction behaviors between users and items to obtain more accurate embedding of user items. Existing social recommendation models tend to take a homogeneous approach to the interaction effects of neighbourhood nodes on target nodes. We propose a social recommendation model based on deep representation learning of hypergraphs. The model aims to learn embedded representations of users and items from multi-type user-item interaction data efficiently. An attention mechanism that integrating time interval information is introduced for computing the preference weights of each item to the user. Messages are generated with time-interval attention weights the inter-user and inter-item hypergraph information is combined to update the messages and aggregate the information from multiple layers. Extensive experiments on two real datasets show that the proposed models are superior to other advanced recommendation methods in improving the recommendation performance.

This paper constructs an interactive offshore wind power resource assessment system based on the 30-year data of ERA5 wind farm from the European Center for Medium-Range Weather Forecasts (ECMWF), combined with the offshore wind power big data storage technology of Apache Cassandra database and the wind energy resource assessment model, and following the current national standards and technical specification documents. In this paper, 6 wind turbines in Changyi wind farm are randomly selected for authenticity, reasonableness and completeness analysis and verification, which verifies the feasibility of constructing wind farms in this sea area, and provides scientific site selection basis and wind resource assessment reference for the construction of offshore wind farms in China.

Urban rail transit train signal system is crucial for ensuring the safe operation of railways. To address this issue, an urban rail transit equipment board Light Emitting Diode (LED) indicators detection system based on image recognition technology is proposed. This study realizes the status and color recognition of equipment board LED indicators based on the You Only Look Once (YOLO) architecture. After training and analyzing the algorithm model, the detection system can efficiently and accurately detect equipment LED indicators while ensuring fast processing speed. The results demonstrate that the urban rail transit equipment LED indicators detection system is highly efficient and has a low omission rate. It significantly reduces the costs associated with manual inspections and substantially improves the intelligent Operations and Maintenance (O&M) capabilities of urban rail transit equipment.

With the further advancement of industrial technology, the data generated by sensors is gradually becoming more complex. Deep learning approaches have made notable strides in the domain of anomaly detection, especially when it comes to high-dimensional datasets. However, most existing methods neglect the complex relationships between feature dimensions and the dependencies between multiple time steps, resulting in insufficient extraction of key feature information. In response to this problem, our paper introduces a model for detecting anomalies in multivariate time series, utilizing Sparse Autoencoder and Dynamic Graph Attention Network (SAE-GATv2). The outcomes of experiments conducted on two actual sensor datasets reveal that our suggested approach is capable of identifying anomalies with greater precision and showcases superior overall performance in comparison to the baseline methods.

Protecting privacy in image data has become increasingly crucial with the rise of digital media and information sharing. Effective classification of privacy-sensitive images is essential for implementing robust privacy protection systems. Based on the nature and sensitivity of the image data, this paper proposes a new classification method of privacy images on the basis of reviewing the current privacy data. For each class, this paper discusses the detection and corresponding data processing techniques based on traditional techniques (image blurring, image downscaling) and deep learning methods, to enhance the overall privacy protection system, Furthermore, this paper describes the challenges of this topic in diversity, complexity, law, and ethics.

To address the limitations of PointNet's single-scale input, which results in incomplete feature extraction, neglect of lowand mid-level point cloud features, and feature loss caused by max pooling, an improved multi-scale feature extraction network model for point cloud classification based on PointNet is proposed. Farthest point sampling is employed to obtain point cloud inputs at different resolutions, and T-Net from PointNet is used for point cloud feature alignment. A combined multi-layer perceptron is introduced to replace the original multi-layer perceptron, enabling the extraction of both lowdimensional and medium-dimensional features. Global max pooling and global average pooling are then combined to aggregate these features. Finally, the global features obtained through mixed pooling at different scales are fused to form the final global feature vector, which is fed into a fully connected network for classification to determine the probability of each category. The proposed multi-scale feature classification network was validated on the ModelNet40 dataset. Experimental results demonstrate that an overall classification accuracy of 90.7% is achieved by our model, representing an improvement of 1.5 percentage points over PointNet, and a mean class accuracy of 86.8% is achieved, an improvement of 0.8 percentage points compared to PointNet.

The co-occurrence and clustering analysis of literature data on machining surface reconstruction in WoS and CNKI databases were conducted using CiteSpace, revealing the current distribution and cooperation status of machined surface image reconstruction at the national, institutional, and researcher levels, and sorting out research hotspots and cuttingedge trends.

Digital technology has shown great potential in the reconstruction and future planning of historical and cultural urban scenes through virtual reality (VR) interaction and digital twin technology. This article adopts high-precision 3D data collection technology, combined with virtual reality and digital twin technology, to create high-precision reconstruction of historical and cultural city scenes, and conducts user interaction experience testing and analysis. The significant advantages of integration technology in planning accuracy, user engagement, and environmental impact assessment were verified through comparative experiments. The research in this article indicates that by integrating virtual reality interaction and digital twin technology, the accuracy and visual effects of historical and cultural city scene reconstruction can be significantly improved. The experimental results show that the root mean square error (RMSE) is only 0.69mm, texture distortion is controlled between 3.1% -6%, and user satisfaction is as high as 10 points. These data fully demonstrate the effectiveness of the integration technology. Compared with traditional planning methods, integrated technology has shown significant advantages in improving planning quality, promoting public participation, and supporting sustainable development.

In this paper, an improved model GSD-YOLO built upon YOLOv8n is proposed to improve the precision and efficiency in detecting insulator defects. The model introduces three key improvements, namely GSConv lightweight convolution, small target detection Layer, and dynamic detection head (DyHead), which significantly enhance the computational efficiency and detection performance of the model. Experimental results illustrate that the mean average precision (mAP) of the GSD-YOLO model is improved from 87.5% to 91.2% and the recall rate is improved from 81.9% to 84.5%, which significantly outperforms the existing YOLOv8n model while maintaining a low computational complexity. In addition, the model’s ability to detect small-target defects in intricate backgrounds is significantly improved, with a minimum 2% increase in the mean average precision, which effectively adapts to the demands of complex power system environments and better satisfies the practical requirements of power inspections.

I developed a deep learning-based recommendation system to enhance book recommendation accuracy using three datasets: Books, Ratings, and Users. After data cleaning and preprocessing, including handling missing values and standardizing formats, I used TF-IDF vectorization and cosine similarity for content-based recommendations. Using TensorFlow and Keras, I built a collaborative filtering model with embedding layers, regularization, and dropout to prevent overfitting. The model was trained with the Adam optimizer and evaluated using RMSE. Training involved adjusting hyperparameters to find optimal values. The deep learning model was compared with traditional methods like Singular Value Decomposition (SVD), Item-based Collaborative Filtering, and Alternating Least Squares (ALS), and showed significant improvement in recommendation accuracy. The model effectively handles large-scale data and enhances recommendation quality. In conclusion, my deep learning-based system, along with the SVD, Item-based Collaborative Filtering, and ALS models, significantly improves book recommendation accuracy, offering valuable insights for future research and development in recommendation systems. Finally, I selected the neural network model and further fine-tuned its parameters to achieve the best performance.

The occurrence of shadows in images leads to a damage in image quality and also hinders the performance of downstream visual tasks. Despite significant progress in existing shadow removal methods, they are limited in synthesizing highfrequency signal details, resulting in insufficient synthesis quality. Among them, most of methods based on Generative Adversarial Networks (GANs), while capable of removing shadows, often introduce image artifacts and blurring, particularly underperforming in the processing of high-frequency signals, which impacts the clarity of the image after shadow removal. To address these issues, we propose a high-low frequency guided GAN method for image shadow removal. Concretely, our method decomposes extracted features into multiple frequency components (i.e., low and high frequency) and retains the image's contours and structural information through low-frequency attention skip connections. Meanwhile, high-frequency attention skip connections are employed to alleviate the difficulty for the generator to synthesize details, providing the generator with rich frequency information. Additionally, we also introduce a wavelet-based frequency loss function to reduce the discrepancies between the generated image and the real image in the frequency domain, effectively mitigating the occurrence of artifacts and blurring. Finally, extensive experiments conducted on the ISTD, AISTD, and SRD datasets demonstrate the significant effectiveness of our proposed method, which not only enhances the efficiency of shadow removal but also markedly improves the high-frequency detail quality of the de-shadowed images.

Due to the complex marine environment, low visibility, object occlusion, and limited underwater resources and environment, the task of detecting marine garbage is extremely difficult. In this regard, this paper suggests an underwater garbage detection model YOLO-CDAW utilizing the improved YOLOv8n model. First, to improve the separation between underwater garbage and occluders, the CA module is inserted to the YOLOv8n model’s backbone network. Secondly, integrates the CoordConv module alongside the C2f module of the YOLOv8n model, proposes a new CD_C2f module, then utilize it to switch out some of the backbone natework’s C2f modules. This module enables the model to get full utilization of the position data relating to the object without increasing the model calculation amount and parameter amount, thereby improving the network's feature extraction ability. Finally, to strengthen the model’s capacity for generalization, the CIoU loss function is swapped out with the WIoU loss function. Experimental results show that the upgrated YOLO-CDAW model’s mAP50/% and mAP50:95/% are 2% and 1.8% greater compared to the original model for each, with an inference speed of 3.3ms and only 3.03M parameters, which fulfills the need for real-time underwater garbage detection within the limits of resource and environmental constraints.

Osteoarthritis is a leading cause of chronic disabilities. Automatic multi-tissues segmentation of knee joint can help doctors diagnose Osteoarthritis by segmenting knee joint MRI. However, manually segmenting is time-consuming for an experienced expert. Besides, this task is challenging due to the significant morphological differences and close proximity between various tissues in the knee joint. To achieve fast and accurate segmentation, we propose a novel hybrid architecture named nnCSCFormer (Not-another Cross-shaped Channel Transformer). This architecture integrates spatial and channel attention mechanisms to enable rapid and accurate automatic multi-tissues segmentation. By capturing the relationship between spatial and channel dimensions across the entire feature space, our model effectively extracts multitissues’specific information. Additionally, we introduce skip attention to aid the decoder in better preserving original image details. Experimental results demonstrate the efficacy of our model in simultaneously segmenting six tissues: femur bone, femoral cartilage, tibia bone, tibial cartilage, lateral meniscus and medial meniscus. The proposed method achieves superior segmentation performance compared to alternative methods on low-resolution knee MRI and has significant application value in preoperative planning for surgical navigation.

It has significant application value by using LiDAR point clouds to analyze the safety of the power corridor. However, the current semantic segmentation network for point clouds has unsatisfactory feature extraction ability for typical targets (such as towers, power lines, and trees) in power corridors, and the segmentation accuracy is not high. In this regard, to address the unevenness of point clouds, the Fused Features and Attention Enhancement Network (FFAE-Net) is proposed in this paper, which mainly includes the proposed Contextual Feature Fusion (CFF) module and Neighbor Attention Pooling (NAP) module. In the CFF module, we integrate the geometric features of dense and sparse regions in the neighborhood, combined with the color change relationship, to deepen the network's understanding of point clouds unevenness and enable the network to fully learn effective information of the target. The NAP module introduces the attention mechanism, to enhance the aggregation for different features of point clouds, and improve the segmentation performance by weighting the main features of the targets in the power corridor. In addition, a new loss function is further constructed to address the unevenness problem of the point clouds, which introduces the geometric deviation between the neighborhood center points and standardized center points. Experiments on LiDAR point clouds of the power corridor indicate that FFAE-Net can achieve satisfactory segmentation results for the power corridor, and the mIoU for five categories of targets, namely, power lines, high-voltage towers, buildings, plants and the ground, can reach 93.78%, which is significantly better than several compared networks.

The current human society is entering an era of artificial intelligence that integrates data-driven and knowledge-driven approaches. Knowledge graph, a technique that has gained significant attention in recent years within the field of artificial intelligence, utilizes a graphical structure to represent knowledge and model the relationships between various entities in the world. The essence of constructing a knowledge graph lies in efficiently and accurately uncovering the associations among different entities. Data mining, which involves extracting potentially valuable information and knowledge from large datasets, provides technical support for building knowledge graphs. This paper primarily discusses the application of data mining techniques in constructing knowledge graphs, proposes implementation pathways for these techniques at various stages of knowledge graph construction, and offers support for creating high-quality knowledge graphs.

In the domain of autonomous vehicles, 3D object detection technology plays a crucial role in improving driving safety and efficiency. This study aims to enhance the ability of the PointRCNN model to identify and detect small-scale objects. To address the limitations of existing models in detecting small objects, we propose an innovative multi-feature fusion network architecture. Initially, this architecture utilizes voxel feature extraction technology to simultaneously acquire point cloud features and voxel features, including Voxel and Pillar features. Subsequently, by comprehensively integrating these features, a more comprehensive and enriched feature representation is constructed, thereby significantly enhancing the model's understanding of target spatial structures and the accuracy of small-scale object identification. Lastly, the integrated features are utilized in the SegNeXt semantic segmentation module, enabling more precise object detection. Experimental results indicate that our method improves pedestrian detection accuracy by 3.84% compared to the original PointRCNN model, thereby validating the effectiveness of this approach in detecting small objects such as pedestrians. This study not only offers new technical support for enhancing the safety and efficiency of autonomous driving vehicles but also provides fresh perspectives and solutions for the advancement of the 3D point cloud object detection field.

With the rapid development of Intelligent Transportation Systems (ITS), vehicle type classification, as a key link in Automatic Toll Collection systems (ATC), is of great significance in improving traffic efficiency and reducing economic losses. This study proposes an intelligent vehicle classification system based on deep learning and multi-sensor data fusion to address the accuracy issues existing in vehicle classification methods based on optical sensors (OS) and human observers. The system significantly improves the accuracy and robustness of vehicle classification by combining deep Convolutional Neural Networks (CNN), LiDAR sensors, and machine learning algorithms. We first constructed a large-scale annotated dataset containing multiple vehicle types and complex traffic scenes to improve our model's capability to identify different vehicle characteristics. Next, CNN models based on different architectures were designed to extract global and local features of the vehicle, respectively. In addition, the LiDAR sensor was used to achieve the spatial structure architectures of the vehicle and combined with the output of the CNN model to improve classification performance under occlusion and complex scenes.

In recent years, the phenomenon of colleges and universities leaking students' private information in information disclosure has occurred from time to time, and the rights and interests of students have been infringed, which has also affected the school's information disclosure work. This article clarifies the content of students' privacy information, analyzes the reasons for the disclosure of private information, and explores how to protect students' private information in disclosure, teachers should do something for the question, students should enhance their awareness, and if necessary, take up the weapon of the law to protect their rights.

In order to realize synchronous motion control of multi-axis system, a motion control design scheme based on EtherCAT real-time industrial Ethernet is proposed in this paper. The multi-axis control system architecture of "High performance system onChip (SoC) +Linux real-time operating system +EtherCAT" was built.

The intrusion of foreign objects into the railway catenary seriously affects the safety of railway operation. In response to the low accuracy and poor real-time performance of current image processing algorithms in detecting foreign objects in the catenary, this paper proposes a lightweight semantic segmentation model SCBAM-DeepLab. This model is not limited to a specific type of foreign objects. Based on the semantic segmentation results of catenary's components, it detects the geometric relationship between invading foreign objects and catenary's components, which meets the operational environment requirements of the real catenary. Based on the lightweight network MobileNetV3, SCBAMDeepLab further simplifies the model parameters and improves the speed of foreign object detection according to the requirements of foreign object detection. Secondly, a module called Superior Convolutional Block Attention Module (SCBAM) was designed to enhance semantic information and the relationships between feature layers, improving the segmentation progress of the components. The model innovatively established a foreign object detection algorithm based on geometric features, analyzing the geometric features of each component to determine whether there is foreign object intrusion. The experimental results show that compared to other advanced semantic segmentation models, the mean intersection over union (MIoU) of this model has increased by about 4%, and the latency has decreased by about 60%. In addition, the mean average precision (mAP) of foreign object detection based on geometric features reaches 86%. SCBAM-DeepLab has improved the detection speed and accuracy of foreign object intrusion with unknown types.

Image recognition technology, leveraging the analytical capabilities of deep learning models, has become an advanced non-contact monitoring tool in the livestock and poultry industry, achieving efficient monitoring and analysis of livestock and poultry behavior, thereby improving production efficiency and animal welfare. This article outlines the development and basic principles of image recognition technology, with a focus on the current application of image recognition techniques such as object detection, image segmentation, and pose estimation in the recognition of animal behavior. It also highlights the challenges faced by image recognition technology in the application of animal behavior recognition and proposes optimization strategies. This paper aims to provide a reference for accelerating the efficient and sustainable development of the breeding industry, promoting the deep integration of artificial intelligence and breeding, and driving the continuous development of the domestic modern breeding industry.

Robotic bin picking is still a challenge problem in industrial field. To address the difficulties of picking parts from a cluttered scene in a bin, an optimal grasping part selection strategy is proposed in this paper. First, parts in the bin are identified and segmented by utilizing the Yolov8 deep learning network. And then, the RGB mask image and depth image of the segmented parts are adopted as inputs for obtaining the optimal grasping strategy. According to the depth data of each segmented part, the features such as average and variance of its depth values, as well as the number of peripheral points that their depth values exceed the threshold are extracted. Subsequently, a weight calculating method is designed for choosing the part easiest to grasp. Finally, the most suitable part is selected for grasping according to the strategy. Unlike traditional bin-picking methods, this strategy aims to identify the part that is easiest to grasp in the bin, thereby avoiding issues such as overlapping occlusion and difficulty in recognition. Experimental results demonstrate that the proposed method improves the success rate of part grasping for bin picking.

Optical fiber composite overhead earth wire (OPGW) is a special cable installed on the top of transmission lines, and its icing problem seriously affects the safe and stable operation of transmission lines. Real-time monitoring of the icecovering condition of optical fiber cables and accurate identification of the ice-covering thickness are essential to ensure the safe operation of power grids. However, the existing ice thickness identification methods have the problems of realtime monitoring difficulty and low detection efficiency, therefore, this paper proposes an OPGW ice thickness identification method based on SABO-KELM. Firstly, the temperature and strain data of OPGW fiber optic cables under different ice cover thicknesses are collected by BOTDR equipment. Then, the ice thickness features in the ice cover data are extracted using the variable modal decomposition (VMD). Finally, SABO is used to optimize the parameter selection of KELM to construct an OPGW ice thickness identification model based on SABO-KELM. Experiments show that the present model achieves a precision and recall of 93.75% and 93.0% compared to other ice thickness recognition models

Human posture estimation is an important content in computer vision, which is the research basis of action recognition, human-computer interaction and other directions, and the traditional manual feature approach has certain problems in terms of accuracy, robustness and computation. In recent years, deep learning-based pose estimation methods have achieved high performance in human pose estimation. According to the number of people, this paper classifies deep learning-based 2D human pose estimation methods into single-person pose estimation methods and multi-person pose estimation methods, and summarizes and analyzes each method. A brief overview of the current mainstream datasets and evaluation metrics is provided. Finally, the development of 2D human pose estimation methods is summarized and possible future trends are listed.

Aiming at the problem of low efficiency and low accuracy of stainless steel welded pipe weld surface defect detection in industrial scenarios, a welded pipe defect detection algorithm based on the improved YOLOv5s is proposed. First, the C3 module in YOLOv5s is replaced with the C2f module, which has fewer parameters and is more sensitive to small targets, to obtain richer gradient information and enhance the feature extraction capability. Secondly, the SimAM parameter-free attention mechanism is introduced to strengthen the model's simultaneous attention to channel and spatial information, which gives higher priority to defective features without increasing any number of parameters, improves the accuracy of the model, and improves the detection performance of the model in various scenarios. The experimental findings demonstrate that the improved algorithm, compared with the traditional YOLOv5s algorithm, improves the mAP by 4.4%,the frame rate FPS reaches 140F/s, and the number of parameters is reduced by 18%; it is easier to be deployed on the inspection equipment in the production environment while meeting the need for rapid and precise detection. The approach offers robust support for detecting defects in stainless steel welded pipe, and provides a reference for inspection intelligence in on-line production of welded pipe.

Traffic sign recognition is an important task in driverless technology. Because the traffic signs are located in a complex environment and the target scale changes greatly, it brings a burden to the network optimization. In reality, it is usually difficult for common models to achieve good detection accuracy while ensuring the detection speed. In order to solve this problem, we introduce the structural re-parameterization technique on the basis of YOLOv5, try to take advantage of the ad-vantages of high performance of multi-branch model and fast reasoning of single-channel model at the same time, and embed the effective channel attachment (ECA) module on the prediction head to find the attention area in complex scenes. Extensive experiments on Tsinghua-Tent 100K (TT100K) dataset show that our improved model can give consideration to recognition speed and accuracy with practical application value

Aiming at the problem that the current pavement crack detection methods have less research on the repair trajectory planning for subsequent cracks, a raod crack detection and repair trajectory planning approach based on image processing is proposed. Firstly, the preprocessing method based on augmented nonlinear diffusion denoising is used to remove pretzel noise points; then, the FCM algorithm based on adaptive weighted spatial information is used to segment the crack features; then, the connectivity domain area filtering is used to remove the large noise points; finally, the coordinate point extraction and polynomial least squares are used to carry out the repair trajectory planning, and combined with the pavement crack detection and repair prototype simulation of the actual scene testing, to validate the reliability of the proposed approach.

The detection and identification of tool marks have always been an important research content in the field of forensic science and criminal science and technology. Traditional methods of tool trace inspection and identification usually rely on the experience and subjective judgment of experts. It is likely to lead to the analysis of judgment bias in the comparison process of the two differences. Therefor the digital extraction and analysis of linear trace has become a hot research topic for improve the reliability of linear trace analysis. In this paper, the single-point laser scanning device is used to extract linear traces on the surface of Aluminum wires. The validity and accuracy of linear trace alignment test are investigated. The matching results of linear traces on 10 fracture surfaces of varying diameters are calculated over a span of 15 seconds, with a success rate exceeding 80%. The results show that linear trace extraction can be used to identify and analyze tool traces.

Landmark detection is a crucial task that focuses on identifying specific landmarks representing distinct object features within an image. However, existing methods primarily adopt Convolutional Neural Networks as the backbone for feature extraction, which struggle to capture global contextual information. Due to the structural dependencies of landmarks, these methods fail to effectively utilize these relationships, leading to lower detection accuracy. To overcome these limitations, we propose a novel framework, CTBN, which combines CNNs, Transformer, and Convolutional Block Attention Modules to enhance feature representation for landmark detection. In our framework, CNNs extract fine-grained local features, Transformer captures global context, and CBAM adaptively refinesfeatures by focusing on important regions and channels. We validate our performance on multiple datasets, including CelebA, AFLW and Deepfashion. The experiments demonstrate that our method achieves superior performance, significantly improves the accuracy of landmark detection.

In the current wireless communication system, WiFi-based wireless sensing is considered as one of the most promising wireless sensing technologies with its extensive deployment and many applications. However, the existing models and methods have some problems, such as low recognition accuracy and poor robustness. To solve the problem that the extraction of human activity information as features are affected by environmental noise and human body posture, and does not take into account the different particle size effects of different actions in different parts of the human body. This paper proposes an action component extraction method based on CSI signal decomposition to extract dynamic components independent of environment and body posture. Finally, a fusion network based on two-stream feature extraction is designed to integrate multi-scale spatial features and temporal features into multi-space temporal features. A large number of experiments are carried out to evaluate the performance of the proposed method, and the experimental results show that the proposed method has a high recognition accuracy.

In computer-aided diagnosis, abdominal multi-organ segmentation is essential, and it has significant research implications. However, the ambiguous boundaries, complex backgrounds, and variable shapes and sizes of abdominal multi-organs make this task extremely challenging. To this goal, an abdominal multi-organ segmentation network called EJTransUNet—a fused convolutional and multi-scale TransUNet network—is proposed. Three components make up the network: a jump connection, a decoder, and an encoder. Regarding the input picture, firstly, a global local enhancement module (DCE) is designed in the hybrid encoder CNN-Transformer module, and the hollow convolution is introduced in the last layer of the Transformer module, which enhances the extraction of global local contextual information, so that the model strengthens the local correlation of the features while establishing the long-distance dependency; secondly, a jump connection stage is designed in the Multi-scale fusion module MSC, which utilizes the CBAM attention mechanism module to capture the correlation between features at different levels and focus on features at different spatial locations to improve the accuracy of target localization; and finally in the decoder output image. Using the Synapse dataset, the model's performance is assessed, yielding an average Dice Similarity Coefficient (DSC) of 80.34%. The experimental findings demonstrate that the suggested approach performs better overall than multiple comparison networks and yields superior segmentation outcomes for organs with varying sizes and shapes.

In recent years, with the rapid development of sensing technology, human subject communication and perception have gradually attracted the attention of researchers. Based on this, we construct Wi-MP in this paper, a method of using CSI data to count people in a room. In this method, two new features based on subcarrier amplitude data extremum difference (DEV) and pyramid-based extremum (m_pool) are proposed for the first time, and then four common statistical features are combined to form a new feature set. Then, the feature collection is sent to the support vector machine (SVM) for classification and recognition. Finally, the classification results are obtained, and the results show that the recognition accuracy of the crowd counting method is 97%, which is higher than that of other crowd counting models.

With the rapid development of wireless sensing technologies, WiFi channel state information (CSI) has become a powerful tool for monitoring human vital signs, such as breathing and body movement. However, estimating respiration rate at rest is often constrained by body motion and non-target signal interference. In this study, we address this challenge by proposing an innovative joint oriented vector technique that accurately distinguishes between target and non-target signals and effectively eliminates the interference of non-target signals. An adaptive sliding window is used to compute the differential standard deviation of the CSI amplitude information, which is combined with a signal-to-noise ratio-optimised subcarrier selection strategy and a PW-Hampel filter to significantly suppress the interference caused by body motion and hardware noise to the respiratory monitoring signals, and an improved peak detection algorithm is used to accurately identify peaks in the respiratory signals. The experimental results show that the proposed method achieves a high accuracy of 98.90% under body movement conditions, which significantly exceeds the level of existing technology.

In this paper, an innovative ring-coded marker point design and its recognition method are proposed to effectively solve the difficulties in localization and detection due to wear and tear or occlusion in complex environments. By adding solid circles as starting characters above and below the marker points, the localization accuracy and recognition stability are significantly improved. Experiments prove that the recognition accuracy of the method is 98.9% even when the occlusion rate is not more than 50%, which meets the requirements of high-precision measurement. The method achieves effective image processing through global binarization and adaptive determination of threshold by Otsu algorithm, and accurately locates the coded marker points through boundary extraction, parameter fitting and other techniques. In addition, this paper discusses the application of coded marker points in photogrammetry, emphasizing the important role of coded marker points in fast and accurate measurement of 3D close-up photography as well as the conversion of images from 2D to 3D. This research not only improves the accuracy and automation of visual measurement techniques, but also provides a new solution for precise positioning and accurate identification in complex environments, which has a wide range of application prospects.

In the task of ship target detection, due to the complex image background and more irrelevant interference, the difficulty of ship target detection is increased. A lightweight and efficient ship object detection algorithm YOLO-MBS based on improved Yolov5 is proposed. Firstly, the backbone network adopts MobileNetv3, which uses lightweight deep separable convolution and residual block structures and a new activation function h-swish to effectively reduce the number of parameters of the network model and improve the calculation speed. Then, the E-IOU loss function is introduced to improve the regression accuracy and convergence speed of the prediction box. Secondly, in order to allow the machine to select the region of interest and extract effective features more fully, the SimAM non-parametric attention mechanism is added. Finally, the BiFPN feature pyramid network structure is used to make the features of different scales better integrated and utilized. The experimental results regarding the Seaships dataset indicate that the volume of the improved algorithm model is compressed to 21.8% of the original YOLOv5 model, the training time is shortened by 46.62 %, the parameter quantity is compressed to 18.5 % of the original model, the calculation amount is reduced to 16.45 % of the original model, and the average accuracy is improved to 98.8 %. The experimental results demonstrate that the improved algorithm significantly reduces model parameters, thereby shortening both training and inference times, while maintaining high-precision detection performance.

Pest detection is an essential early alert mechanism within agriculture. Nonetheless, the datasets related to agricultural pests pose numerous challenges, such as a variety of species, minuscule sizes, dense populations, and highly similar appearances, all of which substantially raise the difficulty of both detecting and managing pests. To tackle these issues effectively, this paper enhances YOLOv8s by integrating BiFormer attention mechanism in the Neck layer to capture the associations and dependencies between the features more efficiently. These improvements not only boost the precision of detection but also significantly reduce the models footprint, making it better suited for real-world applications. Testing on the Pest24 dataset demonstrates that our model achieves superior performance, with a detection accuracy of 39.6 AP, surpassing YOLOv8n by 4.7AP, and increases computational efficiency by about 13.5%.

Pulmonary rehabilitation training is one of the important therapeutic measures for patients with chronic obstructive pulmonary disease (COPD) during the clinical remission period. It aims to improve lung function through a series of respiratory muscle training exercises. Common respiratory muscle training includes pursed-lip breathing and diaphragmatic breathing, which can effectively improve lung ventilation efficiency. This paper innovatively applies RFID technology to pulmonary function breathing training, proposing the use of Gramian Angular Field (GAF) and Markov Transition Field (MTF) to convert one-dimensional RSSI breathing signal data into multi-channel image data, more comprehensively reflecting the dynamic changes and characteristics of breathing signals. By classifying respiratory muscle actions through the Mnasnet network model, and inspired by Mnasnet, we propose a Mnasnet-3Block-SVM network model with fewer parameters and shorter inference time. This model achieves a classification accuracy of 96.59% for respiratory actions, an improvement of 2.09% over the baseline model.

Begin Rice, as the largest food crop grown in China, has shown a high incidence of rice blast disease, which seriously threatens the sustainability of rice production. A UAV multispectral remote sensing platform is used to realize the identification and monitoring of the incidence of rice blast disease and early warning at the field level, which assists variable fertilizer and drug application by plant protection UAVs. This research is based on a rice blast experiment in a large field in Wanchang Town, Yongji County, Jilin City, based on a malefactor, using a UAV multispectral remote sensing platform to obtain images of two fertility stages, namely, the nodulation stage(NS) and the spike stage(SS) of rice, and using six different vegetation index parameters as inputs to establish a DBO-XGBoost classification model, with an accuracy rate of more than 90% for both fertility stages.

In aerial photography and remote sensing imagery, among other domains, a proliferation of small objects is frequently encountered, which, due to their homogenous and sparse textural characteristics, pose formidable challenges to existing object detection algorithms, leading to suboptimal performance in their detection. To tackle these issues, this paper presents an augmented model, MSCF-YOLO, grounded in YOLOv8, with a focus on enhancing multi-scale features and integrating object contextual information. This is achieved through the incorporation of Poly Kernel Inception Blocks and Channel Prior Convolutional Attention mechanisms, thereby enriching the model's feature representation capacity. Additionally, the introduction of the Powerful-IOU guided small object anchoring strategy facilitates regression towards more favorable paths, thereby augmenting the model's learning proficiency pertaining to small objects. Empirical assessments conducted on the test set of the VisDrone2019-DET dataset revealed that the proposed MSCF-YOLO model achieves a mean Average Precision (mAP) of 42.2%, surpassing its baseline counterpart, conclusively affirming the efficacy of the proposed enhancements.

With the continuous development and expansion of power systems, the safe operation of transmission lines has become particularly crucial. In response to the shortcomings of traditional methods, which require manual adjustment of parameters and features, this paper proposes a deep learning-based line ice extraction algorithm, IDR-Net. IDR-Net includes a multiscale ice feature extraction module that can extract features of ice at various scales on the line, facilitating the extraction of features from ice-covered targets of different sizes. IDR-Net's DualUP tandem dual upsampling module can prevent the checkerboard effect commonly associated with traditional upsampling modules, thus enhancing the capability of feature extraction and the precision of image edge segmentation during upsampling. Additionally, IDR-Net's residual ice feature extraction module helps avoid the issues of gradient explosion and disappearance, allowing the network to retain more feature information. After experimenting on a self-constructed dataset, the results show that the improved network achieved a mean Intersection over Union (mIoU) of 81.55%, an increase of 4.13 percentage points, demonstrating that the improved network can effectively enhance the accuracy of remote sensing image segmentation.

Rice production is the key factor to ensure food security. Rice varieties classification is an important means to improve rice quality. Considering the problem that traditional machine learning methods lack analytic solutions and can not adapt parameters. This paper implements the ELM method based on two-stage weights. In the initial stage, the weights of the output layer can be calculated according to the Moore-Penrose generalized inverse. Then regular function balances the expectation error and model complexity. In the model optimized stage, we calculated the model weights of the new function according to the predictive performance. These strategies help to reduce the fitting error of sequence fragments. Numerical experiment results are proved at the University of California Irvine data set, the results show that the proposed method not only has better generalization performance, but also guarantees the prediction effect of contrast method for different scenes.

Micro-expressions (MEs) are unintentional, brief facial expressions with the potential to reveal hidden or suppressed emotions, significantly contributing to the study of human emotions and psychological states. Despite this, the brief duration and subtlety of MEs make micro-expression recognition (MER) a difficult task. This paper introduces a domain adaptation based multi-scale feature pyramid perception framework for MER, namely domain adaptation multi-scale feature pyramid perception network (DAMFPP-Net), extracting common feature representations between MEs and macroexpressions (MaEs) and transferring them to MER task. Specifically, the network extracts multi-scale features of both MEs and MaEs by multi-scale feature pyramids construction and perception module. Then, domain adaptation is utilized to reduce the differences in feature distributions between the two, thus obtaining a shared emotional representation. Finally, the common features are transferred to MER task to adapt to the unique characteristics of MEs. Our approach demonstrates state-of-the-art performance through extensive experiments on three public micro-expression datasets.

This study aims to develop a video scene recognition model that integrates multimodal information to improve the accuracy and robustness of video scene understanding. The model combines visual, audio, and text data, and achieves effective fusion of multimodal information through efficient data preprocessing and feature extraction mechanisms. Experimental verification shows that the model exhibits excellent recognition performance in various scenarios, especially in complex scene processing, outperforming traditional single modal methods.

There is often a phenomenon of inconsistency between the actual personnel and the staffing of military units at all levels, which leads to the failure to maximize the benefits of military human resources. In order to maximize the matching between staffing and serving officers, it is crucial to grasp the matching degree of staffing in time and accurately. Therefore, with the aim to realize multi-index analysis and layer-by-layer synthetic evaluation, an evaluation model is established in this paper for the matching degree of staffing and serving officers based on the Analytic Hierarchy Process (AHP), forward transfer network, fuzzy statistics, and other methods. The simulation results prove the effectiveness and reliability of the method, which can improve the accuracy of the evaluation, reduce the amount of data processing, and provide some scientific basis for the rational allocation of military personnel.

As the field of self-driving cars accelerates, the accurate interpretation of traffic officers' signals has become a pivotal technological advancement. Current research has explored various feature representation-based methods for traffic police gesture recognition, including RGB frames, optical flow, human skeletons, and point clouds. However, the existing methods still face challenges in spatio-temporal information fusion and feature extraction. Specifically, the skeletonbased method, noted for its focus on actions and compactness, could benefit from improvements in processing speed and robustness. In response to these challenges, the present study proposes a SlowFast network-based approach for traffic police gesture recognition (SF-TPGR). The method uses 3D heatmaps as inputs to convey the relative positional information of joint points through heatmap intensity changes, thereby improving model accuracy and speed by reducing the redundancy of heatmaps for lower limb joints. The SlowFast network architecture is then used for traffic police gesture recognition, effectively modeling rapidly changing actions such as going straight and lane changing with high frame refresh rates. The comparative experiment conducted on the Chinese traffic police gesture dataset shows that the accuracy of this method reaches 98.10%, and the inference time is reduced to 79 ms. Although its accuracy may be slightly lower than some top-performing methods in certain cases, its capability to rapidly extract human movements is a significant advantage.

The intermediate joint of the three-core winding package is a critical component in distribution network cable lines, and its standardization and consistency directly impact the safety and reliability of cable operations. However, due to the complex structure, high construction difficulty, and inconsistent acceptance standards of the intermediate joint, its quality can vary significantly, leading to potential hazards such as tip discharge, partial discharge, and overheating. To address these challenges, this paper introduces a deep neural network-based identification technology for the intermediate joints in three-core winding packages, utilizing the YOLO algorithm. By intelligently analyzing and assessing key characteristics such as shape, size, and position, our method enables automated identification and evaluation, thereby providing a scientific foundation and technical support for improving the quality control and reliability of cable joint production. This innovative approach significantly enhances the accuracy and efficiency of defect detection, ultimately contributing to enhanced safety and operational reliability in the field.

The traditional fire warning system with a single threshold value is widely used. However, due to the development process of fire, the variation of parameters in each stage is quite different, so it is difficult to achieve rapid and accurate fire warning with a single parameter. Aiming at the problems existing in the traditional fire alarm system, a BP neural network prediction algorithm based on improved artificial fish swarm optimization is proposed. Combined with multi-data fusion technology, the characteristic temperature, CO concentration and smoke concentration in the early stage of fire are used as inputs, and the three stages of fire development are used as output of BP neural network for fire prediction. On the basis of the standard artificial fish swarm optimization BP neural network, chaos algorithm is used to improve the artificial fish swarm, optimize the search ability of artificial fish swarm algorithm, accelerate the convergence speed of neural network, and output the expected results. The experimental results of MATLAB simulation software show that compared with the standard artificial fish swarm optimization BP neural network, the convergence accuracy is improved, which proves the superiority of this method.