Underwater environments can be used to explore new resources that can be employed in the fields of medical science and energy resources. Humans are dependent on the valuable resources that exist beneath Earth’s surface. Underwater exploration requires enhanced images that are obtained using enhancement methods. So, it is important that underwater image enhancement (UIE) methods work well in terms of performance and accuracy. As a result, research in UIE has increased in the past few years. An extensive survey is conducted on existing UIE methods along with their broad classification, underwater datasets, and evaluation metrics, respectively. The experimental analysis is conducted to compare the existing UIE methods in terms of qualitative and quantitative evaluation metrics. The real-world applications and future scope of existing enhancement methods are highlighted and discussed.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

AB, KK, HJA, and MK did not agree with the retraction. MDA either did not respond directly or could not be reached.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

HZ, MN, XC, JW, YZ, and CL either did not respond directly or could not be reached.

With the rapid development of automation technology, unmanned aerial vehicles (UAV) have been widely used to assist troops in long-distance combat missions. However, considering that the UAV needs to keep flying and maintain a certain distance from the target, face detection using UAV usually achieves poor results, which hinders the application of the face recognition technology in UAV systems. We proposed an improved Mish-L2-multitask convolutional neural network (MTCNN) model based on MTCNN model to further improve the accuracy of small-size face detection. First, the maximum pooling layer of the P-Net CNN was removed. Second, a regularization term was added to the crossentropy loss function. Finally, the activation functions exerted in three subneural networks of MTCNN model were replaced with Mish activation functions. The result shows that the proposed Mish-L2-MTCNN model could improve the accuracy of small-size face detection efficiently under the UAV view. Comparing with the result obtained from the original MTCNN model, the accuracy was improved by 3.62%, and the missing rate was evidently reduced. This work can provide methodological guidance for the development of a UAV-based face recognition system in the real airborne scenario and guarantee the validity and efficiency of the system. The study can also guide further research concerning the detection effect in the cases of side face and half mask.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

ASM, AK, and NK either did not respond directly or could not be reached.

Currently, autonomous vehicles and intelligent robots are used in an extensive set of industrial applications. However, in the recent past, the frequency with which small obstacles arise on highways has risen dramatically, which will lead to severe incidents on highways. Hence, small obstacle detection is critical for increasing the efficiency of autonomous vehicles to enable their safe navigation to avoid road accidents. On the other hand, these small-size obstacles are of varying sizes, shapes, and colors and are difficult to be detected under low lighting and illumination conditions. Existing research studies suggest deep-learning-assisted semantic segmentation models; however, an optimal model along with improved performance is of greater necessity. From this perspective, we have suggested transfer learning-based approaches using state-of-the-art semantic segmentation models namely UNet++, PSPNet, PANNet, LinkNet, and DeepLabV3+ for the detection of small-size obstacles under strict lighting and illumination conditions. Furthermore, we used images of road scenes with extremely small-size obstacles, which were neglected by past research studies, since these obstacles were not included in the databases that they were using in their findings and may have failed to effectively address the problem of obstacle detection. Moreover, for faster and better convergence, we have modified the backbone architectures of these models with Residual-Network (ResNet)-18 and ResNet-34 trained on ImageNet weights. It is observed that DeepLabV3 + ResNet-18 as backbone architecture shows the highest results with a mean intersection-over-union score of 64% along with a 95% value of accuracy.

Empowering retail service robots with empathy is one of the current research hotspots in the field of artificial intelligence. Identifying consumer emotions, understanding the changes in shopping interests, and developing appropriate sales strategies is a challenging task for retail service robots. We investigate the feasibility of using computer vision methods for empowering robots with empathy by examining the correlation between consumer emotion and levels of shopping interest. To this end, we construct the first video database of consumer sentiment changes in a business context and propose a deep learning method that uses multimodal information to infer consumers’ shopping intentions, and conduct preliminary experimental validation on this database. The experimental results show that the proposed method is 7% and 10% more accurate than manual assessment (n  =  40) in identifying consumer emotions and predicting consumer shopping interest levels, respectively. Thus, the proposed method is valid and effective. We anticipate that the results of this study will have considerable implications for human–computer interaction research in service robots.

In the field of computer vision, face recognition has become a trending research topic, and is widely used in the area of network security. The transmission of data over a network via a cloud server exposes the information to security risks and privacy attacks, meaning that the use of an encryption algorithm is essential. Face recognition algorithms in robotics applications have become cumbersome in terms of the computation speed needed to recognize the image. Since this is a built-in programming function in the robotics board, it will limit the speed and security of data storage. To overcome this issue, a cloud server is utilized as this can improve the processing speed, throughput, efficiency, and robustness of face recognition. Storing images in the cloud server in a secure way requires that the image be encrypted. To achieve this, we propose a hybrid encryption technique based on bit slicing and a discrete Fourier transform (DFT), and develop a secure robotic face image recognition scheme using a MobileFaceNet-CNN model (BS-DFT-MobileFaceNet). The percentages of improvement in terms of accuracy over LeNet, VGG16Net, Alexnet, GoogLeNet, ResNet, MobileFaceNet including raw input image were 17.5%, 7.36%, 22.41%, 22.62%, 8.65%, and 288.89% for raw input images and encrypted images using Genetic Algorithm (GA), DNA Algorithm, Bit slicing, AES, DFT, and BS-DFT encryption algorithms respectively.

Robots have a wide range of computer capacities, and executing powerful computational programs on them might be difficult due to restricted internal processing, memory, and energy. Similarly, cloud computing enables on-demand computation, so integrating robotics and cloud computing can help robots solve limitations. The key to successfully offloading jobs is an operational solution that would not underutilize the robot’s natural processing capacity and act authorized based on important costing criteria like delay and CPU resources. Applications are offloaded from robots based on the Markovian decision process. The Markovian decision helps to identify the resources in the cloud network based on probability. A deep reinforcement learning-based deep Q-network (DQN) technique selects resources in the cloud network. Further data are offloaded in cloud storage. The state-space is built on the notion that the size of the input information has a serious influence on the software’s processing time. The suggested technique is constructed as a repetitive work issue with a distinct space domain, in which we take a different activity at every successive stage using the resulting result to train the DQN to get the most prizes. A navigation testbed was created and implemented to validate the suggested method. The proposed method minimizes the cost of communications between clouds and also it minimizes the latency of the application. It increases the accuracy level by 85%, which is higher when compared with existing methods.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

The authors agree with this decision.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

WW, LQ, and JJ either did not respond directly or could not be reached.

Illegal short-term vehicle stalls are frequent in developing countries, which disturb smooth traffic flow and results in congestion, road blockage, and even lead to accidents. Short-term stalls span a few seconds—e.g., a vehicle stopping to make way for people crossing the road. However, traditional stall detection techniques detect stalls that span 10 s or more. Moreover, existing techniques, such as background subtraction algorithms, are vulnerable to illumination changes and indicate too many false stalls with smaller stopping criteria. Furthermore, for violation indication, post-stall analysis is required to filter out the legitimate stalls where a vehicle stops for some valid reason. We present a two-fold methodology to detect illegally parked vehicles. An object identification-based technique is employed over stationary images to detect stall. Furthermore, the legitimacy of the stall is determined by performing pedestrian detection and its position analysis. We tested our proposed approach using publicly available advanced video and signal based surveillance 2007 PV dataset and ViSOR dataset. The technique successfully detects stalls with a recall of 0.96, after weaving the stopping criterion from the ground truth.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

SHC disagrees with the retraction. MSA, MAA, and SS either did not respond or could not be reached.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

HL and BX either did not respond directly or could not be reached.

Recent technological development in the computer vision field has provided opportunities for agriculture applications. The field of computer vision allows a machine to “see,” providing an alternative to using the human eye to identify, measure, and track a target in image processing. Computer vision technology uses a sensor, camera, and computer. The carrot is one of the most important vegetables in the world, and a grading scale for carrots will improve market competitiveness. In marketing and carrot processing, carrot grading plays a major part. Existing approaches using the traditional carrot grading system, which requires manual involvement, are inefficient and labor-intensive. The computer vision feature parameters of the carrot, such as its length, average diameter, maximum diameter, aspect ratio, perimeter, and area, are extracted, along with RGB (Red, Green, Blue) and HSV (Hue, Saturation, Value) color parameters. These 12 computer vision parameters are given as input to a recognition model using the proposed optimized cascaded graph convolutional neural network with Bayesian optimization. The statistical results of the experiments performed with this model show that the proposed recognition system with computer vision feature parameters can grade carrots accurately.

We investigate the loop-closure detection problem for a visual simultaneous localization and mapping (SLAM) system. Generally, a deep neural network is used to solve loop-closure detection problems. However, this method is time-consuming and occupies a large memory in practice, primarily owing to the high-dimensional nature of the data involved. Herein, we combine a pretrained convolutional neural network (CNN) model with a convolutional autoencoder to acquire a low-dimensional CNN vector as an image representation. Subsequently, the loop-closure problem is simulated by the single-image nearest-neighbor search method based on the cosine distance of the low-dimensional CNN vectors. In addition, to solve the perceptual aliasing problem in loop-closure detection, we empirically propose an efficient method for eliminating false positive loop closures by appropriately decomposing and reconstructing the similarity matrix. Furthermore, we verify through experiments that our method can effectively reduce the time and memory costs of loop-closure detection tasks while ensuring detection accuracy. Therefore, this method can provide a feasible solution to improve the real-time performance and robustness of visual SLAM systems.

There are many difficulties in cow face recognition, including variable individual poses, challenges in collecting individual data, complex image backgrounds, etc. An improved capsule network (CapsNet) was used to solve these problems. First, we combined convolutional and local binary pattern (LBP) texture features with a feature extractor named C-LBP. Then, by utilizing the self-attention module, the feature extraction capability was enhanced. An intermediate capsule layer was added to improve capsules utilization. We tested our model on datasets of cow faces. According to the experimental data, the proposed model improved the structure of the original CapsNet. In the middle of the training, positional data was added, giving the model a higher performance and greater resilience in cow face recognition.

We introduce an approach of semantic segmentation to detect various objects for the mobile robot system “ROSWITHA” (RObot System WITH Autonomy). Developing a semantic segmentation method is a challenging research field in machine learning and computer vision. The semantic segmentation approach is robust compared with the other traditional state-of-the-art methods for understanding the surroundings. Semantic segmentation is a method that presents the most information about the object, such as classification and localization of the object on the image level and the pixel level, thus precisely depicting the shape and position of the object in space. In this work, we experimented with verifying the effectiveness of semantic segmentation when used as an aid to improving the performance of robust indoor navigation tasks. To make the output map of semantic segmentation meaningful, and enhance the model accuracy, points cloud data were extracted from the depth camera, which fuses the data originated from RGB and depth stream to improve the speed and accuracy compared with different machine learning algorithms. We compared our modified approach with the state-of-the-art methods and compared the results when trained with the available dataset NYUv2. Moreover, the model was then trained with the customized indoor dataset 1 (three classes) and dataset 2 (seven classes) to achieve a robust classification of the objects in the dynamic environment of Frankfurt University of Applied Sciences laboratories. The model attains a global accuracy of 98.2%, with a mean intersection over union (mIoU) of 90.9% for dataset 1. For dataset 2, the model achieves a global accuracy of 95.6%, with an mIoU of 72%. Furthermore, the evaluations were performed in our indoor scenario.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of compromised peer review and determined the paper is out of the journal's scope. The Editor and publisher no longer have confidence in the results and conclusions of the article.

SHC disagrees with the retraction. MSA, MAA, and SS either did not respond or could not be reached.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

NL, XG, and PJ either did not respond directly or could not be reached.

Image registration is an electronic imaging technology that matches two or more images acquired at different times or under different conditions. Here, we highlight the application of image registration in monitoring robots (e.g., autonomous vehicles) of coal mines. In image registration processing, the scale-invariant feature transform algorithm and its improved versions achieve poor extraction effects for feature points on smooth-edge curves. In comparison, the speeded-up robust features (SURF) algorithm achieves more stable registration results in various images. To improve the accuracy of the SURF algorithm, it is optimized using the particle swarm optimization (PSO) algorithm and several improved PSOs, but it achieves unstable accuracy. Therefore, we optimize the improved gray wolf algorithm (IGWO) using advanced mathematical models in several computing steps. We add the memory function of both experienced individual and group optimal locations in PSO to update the particle positions. The evaluations performed on different datasets show that the accuracy and range of registration are substantially improved by IGWO and IGWO-PSO; however, the running time is extended. Thus, our improved method is competitive with other state-of-the-art image registration methods.

Nowadays, automated and intelligent solutions are required by different services and applications to satisfy current industrial necessities. Internet of things/robotic devices have been found useful in the collection and processing of data for pattern recognition, analyzing and anticipating incident information, optimization, and provision of better and timely decision-making to enable network providers to measure the quality of services, thereby ensuring efficiency. Telecommunication systems need reliable and fast internet networks for connected devices and other business activities; thus, technologies are being explored to determine efficient ways of planning and installing base stations. Recently, possibilities of deploying unmanned aerial vehicles (UAVs) as base stations had been investigated. Although successes have been recorded, not much work has been done on network planning tools for location optimization of UAV base stations (UAV-BS) in 5G and beyond networks to the best of our knowledge. In this paper, an intelligent network planning tool (iNPT) for location optimization of UAV-BS has been developed with different channel models and network planning parameters. The developed tool has resources to suggest a better positioning of the UAV-BS using the simulated annealing meta-heuristic to optimize their locations. The tool also includes a battery management interface that can calculate the battery dissipation with reference to the received and transmitted power of the UAV-BS and display the estimated time of flight for the UAV-BS. Geographical information, such as the height of a specific region, obtained from a map is considered in the simulations. The results obtained using the COST231-Hata, Hata path loss, log distance, and two-ray ground reflection models have been presented in this paper. Based on the experimental results, the developed simulator can be useful in network planning tasks to improve coverage areas using UAV-BSs. In addition, the proposed tool has a friendly interface that provides geographical images, and the BSs and UAVs optimized locations.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

PU disagrees with the retraction. NR either did not respond or could not be reached.

Polygonal approximation (PA) is a challenging problem and has its wide applications in computer vision, electronic imaging, image classification, image visualization, pattern recognition and image analysis. This paper proposes a stochastic technique-based firefly algorithm (FA) for PA. FA customizes a kind of randomization by searching a set of solutions where PA requires more combinations of approximation to find an optimal solution. The significant feature of the FA is real random sample generation and attractive features. The attractiveness and brightness of the firefly have been used efficiently to solve the approximation problem by detecting the dominant points (DPs) with fewer iterations to produce better results. Thus, this technique achieves the main goal of PA that is a minimum error value with a lesser number of DPs. While compared with other similar algorithms, FA is independent of velocities which are considered as an advantage for this algorithm. Subsequently, the multiswarm nature of FA allows finding multiple optimal solutions concurrently. The experimental results show that the proposed algorithm generates better solutions compared with other algorithms.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

YY, HP, XZ, RZ, and XZ either did not respond directly or could not be reached.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

LL either did not respond directly or could not be reached.

Digital x-ray medical image with human visual characteristics is a kind of medical image formed by x-ray irradiation. Currently, medical image processing algorithm is the main hot topic. With the popularization of medical images in clinical diagnosis and the development of computer technology, the accuracy of clinical diagnosis can be greatly improved using medical image processing technology. For digital x-ray imaging in medical system, it is inevitable that the quality of medical image will be degraded by various factors, so it needs to be processed before they can be used. At the same time, with the development of medical digital x-ray detector technology, the range of medical images used will be much smaller than that of detectors. The medical digital x-ray image data received after detection also need to be processed by an effective image-processing algorithm before display. On the other hand, the research, analysis, and diagnostic use of computer in medical images can appropriately reduce the work intensity of doctors or medical staff, and also play an important role in the discovery of certain diseases. A method of digital x-ray medical image enhancement based on human visual characteristics is studied by analyzing the characteristics of histogram of x-ray image and the maximum entropy fuzzy threshold image enhancement algorithm. The experimental results show that the method can fully recognize the medical x-ray images with incorrect exposure, enhancing the effect of digital x-ray imaging and improving the recognition rate.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

ML did not agree with the retraction.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

RM and LC either did not respond directly or could not be reached.

Performances of the deep learning models for person identification and verification are degraded on low illumination images due to the generated facial embeddings being unable to be matched with the trained good illumination facial embeddings. Person verification on low illumination images is a challenging task. The existing techniques have adopted an approach of enhancing the low illumination images and performing the person verification in the enhanced images. But these techniques have not achieved satisfactory results, because the gamma value is kept constant in the power law intensity transformation function to enhance the images. To obtain better performance, we propose a deep learning-based framework that consists of a contrast enhancement module, called as contrast enhancement network (CENet); person identification; and person verification modules. The CENet is built based on the residual network, which predicts the gamma value based on the illumination of the input image. The predicted value is used to perform gamma correction on the image to improve the brightness difference between the faces and their background, whereas the existing techniques are keeping the gamma value as constant for image enhancement. After performing the image enhancement, the enhanced image is given as input to the person identification module. Then the detected faces are verified by the person verification module. Experimental results show that the proposed framework has achieved an improvement of 3.4% to 13% in person identification and verification accuracy on the extended yale face dataset to the existing methods.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

FC and YF either did not respond directly or could not be reached.

Hyperspectral remote sensing (HRS) is commonly employed for crop and soil mapping. Soil and crop classification mapping acts as major roles in planning and management of sustainable development. Thus, it is necessary to design automated crop and soil prediction techniques using HRS images. The recent developments of machine learning (ML) and deep learning (DL) enable to design of effective classification models for soil and crops. This study effectively introduces improved chimp optimization with deep transfer learning enabled crop and soil classification (ICODTL-CSC) technique using HRS images. The presented ICODTL-SC technique aims to identify and categorize different types of crops and soil. To accomplish this, the ICODTL-CSC technique incorporates Adagrad with EffcientNetB0-based feature extractor for producing a collection of feature vectors. In addition, Improved Chimp Optimization Algorithm with a deep adaptive wavelet network model is applied for soil classification, showing the novelty of our work. The performance validation of the presented method is performed using Tamil Nadu Hills dataset, Pavia University, and Salinas scene datasets. The experimental outcomes indicated that the presented method has exhibited the existing techniques in terms of different performance measures.

Human pose estimation is a fundamental yet challenging task in computer vision. Although many methods have achieved significant improvement, they are still insufficient for the fusion of feature maps at different stages, such as the stacked hourglass network (SHNet). The SHNet is a classic human pose estimation network that extracts multiscale features through stacked multistage downsampling and upsampling operations. We propose a multistage attention mechanism to fuse the multistage feature maps. Furthermore, we apply it in the SHNet to propose a multistage attention network (MANet). In the experiments, we demonstrated the effectiveness of MANet in human pose estimation on the common objects in context dataset and the MPII human pose dataset.

Low-light enhancement is an important technique for improving image quality. This is because low-light enhancement is expected to improve image visibility while maintaining visual naturalness of the image. In recent years, many methods have been researched to enhance low-light images, including histogram-based, fusion-based, and learning-based methods. The most representative and widely used low-light image enhancement method is the so-called Retinex-based method. However, they tend to have many limitations. The limitations of the Retinex-based method are as follows. (1) Due to strong imaging noise or less-effective image decomposition, this results in a large number of artifacts in the enhanced results. (2) Although the first problem can be partially solved by exploring prior information, it often complicates the optimization process. (3) Small-magnitude details are often lost in enhanced results. To overcome these drawbacks, we propose a model called the fractional-order Retinex model. At the same time, Retinex images are decomposed in an effective semidecoupled way. More concretely, the illumination layer T is gradually estimated only with the observed image S based on the proposed variation model, whereas the reflectance layer R is jointly estimated by the intermediate T and S. Experimental results demonstrate the effectiveness of our method.

The purpose of image steganalysis is to detect whether the secret information is hidden in the image. The current advanced adaptive steganography algorithm hides the secret information in the complex area of the image texture. However, it is difficult for the image steganalysis model to capture enough noise residual, which makes the detection ability of the model insufficient. To further improve the detection ability of the spatial image steganalysis model, a U-Net-based auxiliary information generation network is first constructed to increase the size of noise residual in the image so that the model can capture more favorable information. Besides, the spatial and channel attention mechanisms are fused to guide the model to pay more attention to the regions of the image that are globally favorable for steganalysis. To verify the effectiveness of the proposed model, experiments are conducted on the BOSSbase-v1.01 dataset through advanced spatial adaptive steganography algorithms S-UNIWARD and WOW. The experimental results show that the proposed model can improve the detection accuracy by up to 4.5% compared with the current optimal deep learning-based spatial image steganalysis models SR-Net and Siamese-Net.

Though deep neural networks (DNNs) have achieved great success in the computer vision and pattern recognition community, studies show that they are vulnerable to adversarial examples. Adversarial perturbations, usually imperceptible to humans, can be added to benign images to form adversarial examples. Lots of gradient-based methods have been proposed to compute adversarial perturbations. However, these methods compute adversarial perturbations without initialization. A proper initialization of the perturbations is critical to the robustness of adversarial examples. To this end, we propose several adversarial perturbation initialization (API) methods for generating robust adversarial examples. Our work comprehensively analyzes the effect of adversarial perturbations initialization on several white-box attack methods. We conduct experiments on three benchmark datasets: MNIST, Cifar10, and ImageNet. Experimental results show that API improves the attack success rates of adversarial examples. The average recognition accuracy of the target model is reduced by about 3.4% when API is used to generate adversarial examples.

Deep neural networks are vulnerable to backdoor attacks, in which the adversary injects a trigger embedded set into the training process. Inputs marked with the trigger provide incorrect predictions, whereas clean inputs remain unaffected. To erase the latent triggers in models, increasing depth, distribution distillation, and model soup (ID3MS), a defensive solution that operates without prior knowledge of triggers and relies on a small clean set is introduced. The depth of the backdoor model is increased by adding fully connected layer(s) at the penultimate layer. Without a classification layer, the original backdoor and increased depth models are considered as teacher and student, respectively. The student model applies distribution distillation to refit the distribution of the clean set and erase the backdoor triggers. The distilled student model is then recovered with the classification layer and model soup is used to ensemble a collection of models generated by various fine-tuning hyperparameters. The experimental results validate the superior performance of the ID3MS compared with existing defensive techniques against several attacks across datasets.

Bougie location is the crucial step of information extraction in biochips. Many researchers have used algorithms, such as image brightness information or autocovariance, to grid the biochip images. However, these algorithms cannot resolve problems such as the bougie lattice being tilted or there being some bougies with insufficient hybridization. To resolve such problems, we propose a grid location algorithm based on the output of the fifth version of You Only Look Once (YOLOv5) for bougies in lattice biochip images. First, we use the YOLOv5 algorithm to detect and locate the bougies with sufficient hybridization. Second, we build a linear model for bougie location based on the least squares method and use the depth-first search algorithm to detect and locate the bougies with insufficient hybridization. Finally, we use Otsu’s method to extract the gray values of grid areas. The algorithm was tested on a dataset consisting of 100 lattice biochip images. The experimental results show that the average accuracy is 99.76%, and the average detection time is 5.04 s. The algorithm has good robustness and high accuracy and can accurately locate the bougies.

In diversified image style transfer tasks, the goal is to translate an everyday photograph to diverse stylized images conditioned on the style of an artwork. Although several style transfer methods have achieved certain diversity through noise injection, they still have two unsolved problems: (1) relatively limited diversity and (2) significant degradation of quality. In this work, we introduce an effective Gating PatternPyramid block (GPP) to resolve these issues by employing a well-designed multibranch architecture. The main idea is underpinned by a finding that different network architectures capture different style patterns from the same artwork. In addition, the GPP block is compatible with many feed-forward style transfer models and empowers them with the ability of generating diverse stylization results. To the best of our knowledge, this is the first style transfer method that achieves significant diversity without injecting random noise, which provides an inspiring perspective for diverse translation research. We perform qualitative and quantitative experiments, showing the effectiveness and superiority of our method against the state-of-the-art diversified style transfer methods.

Image editing techniques can modify the content of images indiscriminately, which causes a grave threat to the security of society. Hence, the localization of manipulated images is inevitable. A serious challenge for image manipulation detection is the lack of strategies for perceiving global features and refining edges. In this paper, we present a multiscale boundary interaction learning network for image manipulation localization to solve both problems. This network contains an adjacent-scale mutual module to enrich the global perception domain by interactively learning adjacent scale features. It avoids the tremendous noise interference caused by the direct fusion of all scale features. To effectively suppress semantic content segmentation, the boundary pixel disparity module computes interpixel differences at specific angles to enhance boundary artifact recognition between tampered and real regions. The fusion attention module is proposed to combine scale and edge messages, integrating spatial and channel correlations in a compatible way. Extensive experimental results indicate that our proposed method is significantly superior to current state-of-the-art methods on public standard datasets.

Multiple point clouds registration is a prerequisite step for geometry object observation from multiple viewpoints. Most algorithms depend on pose graph or relative poses, and they make the results ambiguous and inconsistent with the original point clouds. However, the existing algorithms based on original data can almost exclusively be applied to simulated data. In this paper, a practical semiautomatic global registration algorithm is proposed. A objective function is proposed; it utilizes correspondences between point cloud pairs directly to describe this problem, making the algorithm easy to interface with other point cloud matching algorithms. Moreover, it can be effectively approximated by semidefinite programming. Since no automatic algorithm can guarantee the correctness of the correspondences, a manual strategy is proposed to filter the correspondences and find the correlation between point clouds. Based on the correlation, irrelevant point clouds can be removed easily in the point cloud set or divide the set into multiple independent parts. Besides, a postprocessing strategy based on the correlation is proposed for refining the results. The qualitative and quantitative experimental results demonstrate that the approach achieves more accurate and robust performance than previous algorithms.

Few-shot classification aims to recognize innovative classes that lack enough labeled examples. Existing few-shot learning methods train a model on the source dataset and transfer knowledge to the target dataset. However, the trained model cannot generalize well to unseen tasks, especially when there are domain shifts. To address the aforementioned limitation, we propose an adaptive fine-tuning strategy to improve the generalizability of the pre-trained model. Our core idea is adaptively transferring target-task-specific knowledge to different target tasks. Specifically, we perform an evolutionary search on the task-specific validation dataset similar to the target dataset. Class hierarchy structure is introduced to seek the closest validation super class of the novel class. Meanwhile, our work takes the evolutionary searched model instead of the pretrained network serving as the initialization of fine-tuning. Extensive experiments proceeded and demonstrate the effectiveness of our adaptive fine-tuning strategy. Furthermore, our proposed method can be easily combined with meta-learning and transfer learning methods to improve their generalizability.

Since intra frame rate control has no prior information to rely on, it is difficult to obtain suitable initial rate control parameters for achieving optimal intra coding. The optimal bit allocation and λ decision are proposed based on a convolutional neural network (CNN) in R-λ intra frame rate control. First, according to the rate distortion (R-D) and rate lambda (R-λ) relationships of the coding tree unit (CTU), a CNN with four outputs is designed to predict the key parameters of R-D and R-λ curves. Second, with the R-D optimization of a frame, the coding optimization equation of the frame λ and the target bit can be built to derive the optimal CTU bit allocation. Last, a suitable CTU λ can be obtained based on the allocated bit and predicted R-λ curve for a CTU to minimize the total distortion of the current frame. The experimental results show that the proposed algorithm can improve the intra frame rate control performance by 0.31 dB with a good rate control accuracy of 4.76%.

The security of sensitive data is critical, and it opens up a wide area of research to find efficient and effective methods to prevent unauthorized access. Our study provides a secure framework for sending visual information over an untrusted channel, such as a social networking site. The proposed framework is a combination of scrambling and encryption techniques. Initially, a hybrid block-wise and pixel-wise scrambling approach is administered to the grayscale image, followed by the Arnold transform, which causes all pixel points to move within the image. Finally, to improve the efficiency of the diffusion process, the asymmetric encryption ElGamal algorithm has been mastered. Peak-signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), number of pixels change rate (NPCR), and unified average changing intensity (UACI) are the metrics utilized to evaluate the efficiency of the proposed scheme. The efficiency of the suggested scheme validates the low-average PSNR value of <9 dB and the SSIM average value of <0.01 for the encrypted images. The NPCR and UACI values achieved in our study are above the threshold values of 99.6% and 33.33%, respectively, exhibiting the strength of the proposed framework.

Graph-based multi-view clustering can effectively reveal the latent cluster structure of multi-view data, however, it remains challenging to construct high-quality graphs by exploring the discriminant information from data. Most of the existing graph-based methods usually construct the graph in the original space or subspace, which may not take the differences of semantic information in different spaces into considerations. Moreover, existing approaches generally measure the relevance of the pair-wise views, which may result in insufficient information exploitation and limited graph representation. To remedy these issues, we proposed an innovative multi-view clustering algorithm called reciprocal consistency graph learning (RCGL), which learns consistency representation by aligning the semantic information from three spaces simultaneously to uncover latent complementarity and consistency among different views. Specifically, we construct the low-dimensional space based on linear projection to preserve essential semantic information. Furthermore, RCGL maps the low-dimension representation from linear space into nonlinear space via Hilbert–Schmidt independence criterion, to explore the nonlinear information hidden in data. Moreover, multiple graphs in kernel space are extended to tensor space for learning a high-order graph, which is helpful to better measure the correlation between multiple views. Finally, the objective function is solved by an alternating optimization scheme. Extensive experiments on five benchmark multi-view datasets demonstrate that RCGL outperforms state-of-the-art baselines markedly.

We propose a dehazing problem model with hybrid regularization and design an effective algorithm to restore the latent image and transmission map simultaneously. In the proposed dehazing problem model, we use the total variation (TV) to regularize the latent image and adopt the hybrid TV and L₀-norm (TV–L₀) regularization to model the transmission map. In the proposed optimization algorithm, we first use the dark channel prior to achieve an initial guess of the global atmospheric light and transmission map. Then we convert the original problem into two subproblems: one aims to update the latent image based on TV regularization, whereas the other estimates the transmission map with hybrid TV–L₀ regularization. Both of the subproblems can be solved efficiently with variable splitting and penalty technology, and the minimizer is reached by alternately solving the two subproblems. Experimental results show that our approach can achieve a high-quality restored image that is comparable to some state-of-the-art methods.

Due to the influence of hazy weather, the images captured often suffer from severe degradation, which will reduce the effectiveness of the devices for computer-related applications. To eliminate the degradation of hazy images, a multiscale image dehazing network based on spatially weighted sparse representation is proposed. The model is mainly divided into two modules: multiscale module (MSM) and detail enhancement module (DEM). The MSM can effectively obtain more feature information in more hazy images. Specifically, the degradation of hazy image is considered in terms of haze distribution and image details. A spatially weighted nonlocal sparse attention module was developed to focus on channel-level features adaptively and to pay more attention to the recovery of dense haze areas in response to the problem of uneven distribution of haze in images. In addition, an independent feature fusion module is proposed to capture better the contextual information, which can also effectively address the traditional fusion, leading to redundant feature extraction and hindering the reconstructive performance of the network. Finally, a DEM is employed by a combination of pyramid pooling and resblock group, which guides the adaptive fusion of latent features, where different details and plausible appearances are integrated. Experimental results on public datasets show that this method is effective and superior in the metrics of structure similarity (SSIM), peak-signal-to-noise ratio, multiscale-SSIM, and root-mean-square error.

Generative adversarial networks (GANs) are generative models based on game theory. Because the relationship between generator and discriminator must be carefully adjusted during the training process, it is difficult to get stable training. Although some solutions are proposed to alleviate this issue, it is still necessary to discuss how to improve the stability of GANs. We propose a GAN we call the discarding fake samples (DFS)-GAN. During the training process, some generated samples are unable to fool the discriminator and provide a relatively invalid gradient for the discriminator. So, in the stabilized discriminator module (SDM), we discard the fake but easily discriminated samples. At the same time, we propose a new loss function, SGAN-gradient penalty 1. We explain the rationale of SDM and our loss function from a Bayesian decision perspective. We inferred the best number of discarded fake samples and verified the selected parameters’ effectiveness by experiments. The Fréchet inception distance (FID) value of DFS-GAN is 14.57 ± 0.19 on Canadian Institute for Advanced Research-10 (CIFAR-10), 20.87 ± 0.33 on CIFAR-100, and 92.42 ± 0.43 on ImageNet, which is lower than that of the current optimal method. Moreover, SDM module can be used in many GANs to decrease the FID value if their loss functions fit.

Raw underwater images are degraded due to light attenuation and particle scattering. It is challenging to obtain authentic reference images. Hence, existing underwater datasets may be short of paired and trainable images. Many data-driven approaches are therefore difficult to implement. To address the above issues, we mainly make two contributions. First, a paired artificial underwater dataset (i.e., PAU-Wild) is constructed, including more than 3500 pairs of images of seven types. PAU-Wild is the first attempt to simulate real-world underwater images with external filters. Second, to enhance underwater images, a convolutional neural network using multistage fusion is proposed. Our model introduces two white balance (WB) priors into the fusion strategy. In addition, a joint loss function is designed to preserve content features and edge features. The experimental results demonstrate that our method is effective and even outperforms several state-of-the-art methods.

Biometrics is widely used in various fields as a means of biological individual identification, and iris recognition is recognized as an efficient, stable, and secure biometric method. However, when the cooperation with users is reduced under complex scenes, the collected iris images are prone to squint or defocus, which causes serious difficulties for iris preprocessing, especially iris location, and segmentation. To solve this problem, we propose a multitask deep learning model based on deep snake. Using the closed-loop prior conditions of the inner and outer boundaries of the iris itself, the active contour method is used to achieve accurate iris location, and semantic segmentation is carried out in the effective region. First, the initial iris location and segmentation results are obtained using the target detection module based on mask regions with convolutional neural networks, then the initial contour is located by poles. The iris boundary is represented by a set of ordered point coordinates through equal spacing sampling. The initial contour is transferred as input to the active contour fitting network based on deep snake to obtain accurate iris location results. Finally, the segmentation results are obtained by the internal and external boundary constraint models. In the segmentation process, efficient channel attention mechanism is introduced to improve the segmentation accuracy and make the network focus on the effective iris region. The SoftPool layer is used to reduce the loss of image detail information caused by pooling, so that the whole network is differentiable. The performance improvements of 3.26%, 3.9%, and 3.45% are obtained on the three squint iris datasets, respectively, which proves the effectiveness of the model in solving the squint iris problem and improves the accuracy and robustness of iris recognition in complex scenes.

Weakly supervised temporal action localization aims to locate the start and end boundaries of action instances and recognize the corresponding categories. Classical methods include random erasure, attention mechanism, and cross-temporal graph relationship modeling. Despite their great progress, there are still two challenges: localization integrity and background interference. Therefore, we propose a framework with self-attention relationship modeling and background suppression to address these issues. First, the input features of background frames are suppressed by the filtering module, which prevents interference from background noise. Second, a self-attention mechanism is designed to model the relationship between different segments in the video, which refines action features to encourage smoother temporal classification scores for completeness localization. Finally, under the guidance of classified loss L_act, the refined segment features and foreground weights are further combined in an attention-weighted pool to achieve video-level prediction. The algorithm is experimentally verified on THUMOS14 and ActivityNet1.2 datasets and compared with other relevant literature, which proves its feasibility and effectiveness.

One-bit compressed sensing (1-bit CS) shows that a sparse signal can exactly be recovered from extremely quantized linear measurements which just catch their signs. While in many applications, real-world signals also exhibit additional structures aside from standard sparsity, such as block-sparsity. We proposed a nonconvex ℓ₂ / ℓ_p ( 0 < p < 1 ) -minimization model. Using the best approximation, covering number, and packing number, we obtain a weak condition to reconstruct block-sparse signals with high probability. And the lower bound of the required number of measurements is lower than some existing 1-bit CS methods. At last, we propose a block adaptive binary iteration thresholding algorithm to recover ℓ₂ / ℓ_p effectively block sparse signals. The algorithm can be used without knowing the sparsity of the underlying signal. Several simulations are conducted to reveal the superiority of our methods to existing approaches, which expose the advantage of 1-bit CS in the reconstruction of block-sparse signals.

Image captioning is a popular research direction at the intersection of machine vision and natural language processing. Most of the existing image captioning methods adopt an encoder–decoder-like structure in which the image is encoded and fed into a decoder to generate a paragraph describing the image content. Although the existing methods have achieved great results in describing natural images, there is still much room for improvement in describing details. We propose the semantic space captioner model to introduce the concept of dense captioning into image captioning using contrastive language-image pretraining as an encoder for text and images. Dense captions are generated for image regions and are used as an extra semantic space for decoding to enhance the final caption. According to the experimental results, our model outperforms existing methods in generalizing image details and is able to generate diverse and meaningful captions. It also performs well on the MSCOCO dataset-related metrics scores.

Some tracking algorithms based on Siamese network have made great progress in similarity learning via features cross-correlation between an object branch and a search branch. However, it is significantly challenging for object tracking in video sequences in terms of target deformation with greatly varying. We propose a Siamese network based on global and local feature matching for object tracking including three phases with the aim of addressing the above issues. In the first phase, obtaining the global similarity matching and local relational mapping similarity of the template branch and the search branch by a selection mechanism of object template-aware features are to reduce the impact of background features on the local matching. In the second phase, introducing correlation matching of the local feature for establishing correspondence among partial-level pixels. Finally, combining the classification and regression results with global matching features and local matching features in a weighted fusion. Extensive experiments are conducted on datasets (OTB-100, LaSOT and GOT-10K) demonstrate that the proposed network enables to achieve superiority compared against the state-of-the-art method and provides an efficient scenario for tackling the issue.

We present a method to estimate the time-to-impact (TTI) from a sequence of images. The method is based on detecting and tracking local extremal points. Their endurance within and between pixels is measured, accumulated, and used to achieve the TTI. This method, which improves on an earlier proposal, is entirely different from the ordinary optical flow technique and allows for fast and low-complex processing. The method is inspired by insects, which have some TTI capability without the possibility to compute high-complex optical flow. The method is further suitable for near-sensor image processing architectures.

Images captured in dust weather have severe yellow and local red, and traditional color correction methods have low robustness. Thus, a dust image restoration method based on channel difference prior (CDP) and a depth residual attention network is proposed in this paper. First, CDP based on statistical observations is defined and used to acquire a dust image dataset. Then, a deep residual network that combines channel-spatial attention is introduced, and channel attention and spatial attention are used to describe important information for every residual block feature. Finally, channel fusion and nonlinear outputs are used to generate clear images. Rich experiments have confirmed the effectiveness and rationality of the proposed method.

End-to-end multiple object tracking (MOT) has been very popular recently in the field of computer vision. However, when performing the inter-frame data association, the majority of algorithms ignore the diversity among trajectories and treat them uniformly. In the joint-detection-and-tracking (JDT) paradigm, there exists an optimization conflict between detection and re-identification due to their different optimization directions. In this paper, we propose a dynamic trajectory quantification (DTQ) strategy for MOT with feature rearrangement. Initially, the DTQ strategy quantifies each trajectory to show its quality, and the score is renewed every frame. To adapt to different situations, we dynamically adopt different updating mechanisms. Additionally, we provide a channel-enhanced feature rearrangement module to alleviate the optimization conflict between subtasks in the JDT paradigm, which obtains more convincing detection results for the DTQ strategy. We evaluate our proposed tracker on several benchmarks, i.e., MOT15, MOT16, MOT17, and MOT20. Extensive experimental results demonstrate that our method achieves competitive results while still maintaining a sizable real-time tracking speed.

Cross-view action recognition plays a pivotal role in some occasions where multiple cameras are required for monitoring. However, it has not performed well yet. One of the main challenges many researchers face is the divergence in viewpoints. This paper focuses on a semisupervised cross-domain feature construction and alignment method for action recognition from cross-view cameras. The proposed framework has four significant steps: First, shape context features are extracted, integrated, and encoded as the pose-level feature representation on a low-dimensional space. Second, key poses are obtained by clustering the pose features and considering the sequence-level feature representation. In the next step, the integrated sequence-level features from different view domains are matched via nonlinear kernelized manifold alignment and brought into a shared space by establishing two different projection matrices to learn a transferable model for cross-view action representation. In the final step, a linear SVM classifier is learned to label the unseen actions from different view domains. Experimental results over benchmark datasets report recognition rates of 90%, 90.91%, and 92.95% on INRIA Xmas Motion Acquisition Sequences, Northwestern-UCLA Multi-View 3D Actions, and multi-view action recognition dataset datasets, respectively, which show that our method achieves superior performance.

Guided by the neural network (NN), NN-based blind image quality assessment (BIQA) methods have improved image quality prediction accuracy. However, the training process of NN will consume a lot of time and training data. To solve the above problems, an efficient NN structure, which is called fast quality assessment network (FQA-Net), is proposed for BIQA. FQA-Net is a very simple NN, which mainly includes a convolution layer, a standard deviation measurement layer, and a regression layer. The convolution kernels in the convolution layer are a set of visual filters with characteristics of visual neurons, which are obtained by training on natural image samples. The output of the convolutional layer is a set of feature maps. Then the standard deviation of each feature map is calculated directly. Finally, a regression model is used for the mapping between the standard deviation values and the quality scores. FQA-Net not only reduces the number of parameters and the output dimensions in the training process, but also prevents network overfitting effectively. The results on seven image databases (i.e., LIVE, CSIQ, TID2013, LIVEMD, KADID-10K, LIVEC, and KonIQ-10K) show that FQA-Net has relatively low computational complexity and high computational speed and accuracy compared with other leading BIQA methods, especially when the number of training samples is very low. Moreover, the FQA-Net can provide stable prediction results across different distortion types.

Due to a growth in information sharing and the use of multiple digital technologies, human lifestyles have been delivered into the virtual world. This digital world has seen the use of images in special fields increase dramatically, especially in healthcare. Healthcare services can be delivered remotely through telemedicine, a well-known method of providing quality healthcare for people worldwide. There is the risk of illegal exploitation of medical data when telemedicine applications involve exposing data over open networks. In particular, medical experts should exercise more caution when sharing a patient’s private information. In the proposed model, reversible data concealment is combined with visual cryptography to provide a secure method of exchanging medical images. A cover image is divided into nonoverlapping secret shares using Hadamard matrix. Secret digital imaging communication in medicine (DICOM) image is encoded by a deep learning model and embedded into secret shares for secure medical image sharing. Finally, the DICOM image is fully reversibly extracted with the cover image. In this case, a visual cryptographic design is used to secure the embedded secret shares. Furthermore, the metrics, such as mean squared error, peak-signal–to-noise ratio, and normalized correlations, are evaluated in the suggested scheme and are compared with various research outcomes to determine the performance of the nominated model.

As one of the development directions of artificial intelligence in the future, few-shot learning has attracted more and more attention in recent years. How to make full use of the information of a small amount of samples is one of the main difficulties in the field of few-shot learning. Most of the research work utilizes the meta-learning mechanism to alleviate the negative impact of insufficient samples on model performance. However, the training between subtasks also makes it difficult for meta-learning models to obtain general feature representations between samples. Therefore, researchers are turning their research perspective to supervised learning, and they have drawn a conclusion that embedding models with good performance are simpler and more effective than complex meta-learning models. Recent research work has also proved the importance of feature representation. Based on the above view points and analysis, we propose a few-shot image classification method, which strengthens the difference of samples from different categories and the similarity of samples from the same category and realizes dual constraints in high-dimensional feature space and low-dimensional feature space. Experimental results on four public datasets demonstrate that the proposed method effectively improves the accuracy of image classification with few-shot learning.

A recent area of study focus is object detecting technologies for marine benthic organisms. A marine benthic object detection algorithm, LSH-DETR, based on DETR, is proposed to address the detection issue in the case of dense distribution of marine benthic small objects. The locality-sensitive hashing attention is used instead of self attention in the transformer to enhance the learning ability of context information. The multilevel feature pyramid networks fusion is introduced to improve the detection performance of small targets. At the same time, repulsion loss is used to optimize the prediction loss of the bounding box to improve the missing detection phenomenon caused by the target stack. According to the findings, LSH-DETR enhanced the mAP of marine benthos to 62.8% and obtained a performance of 13.8% MR when compared to the original DETR. This improvement demonstrates that LSH-DETR is more effective at locating small, dense marine benthic organisms.

Image segmentation is an essential digital image processing problem closely related to computer vision. The limitations of the total variational regular term for the continuous Potts model make the transition of segmentation results over-smooth. In this paper, we propose an improved model with prior conditions to realize the improvement of image segmentation results quality. To extract the structural features of the images, we utilize the simple and practical K-mean clustering (KM) algorithm to set the corresponding volume structure as a prior condition, which is also crucial for the initial label selection of the proposed model. Then, we choose two convex relaxation methods to solve the original nonconvex variational problem. Using these methods, we verify that increasing the constraint of volume structure can maintain slender structures in the original image and achieve a good balance between image segmentation quality and computation. Consequently, we use the unified, convergent primal-dual (PD) algorithm to solve the minimization problem in the proposed model. Extensive experimental comparisons between our method and the pure KM method, the graph-cut method, and the corresponding Potts model without volumetric structure are provided. The segmentation results illustrate that our model performs better in terms of both visualization and evaluation metrics.

Modality-fused representation is an essential and challenging task in multimodal emotion analysis. Previous studies have already yielded remarkable achievements. However, there are two problems: insufficient feature interaction and rough data fusion. To investigate these two challenges more deeply, first, a hybrid architecture, which consists of convolution and a transformer, is proposed to extract local and global features. Second, for extracting more sufficient mutual features from multimodal datasets, our model is comprised of three parts: (1) the interior transformer encoder (TE) aims to extract the intramodality characteristics from the current monomodality; (2) the between TE aims to extract the intermodality feature between two different modalities; and (3) the enhance TE aims to extract the target modality enhance feature from multimodality. Finally, instead of directly fusing features by a linear function, we employ a popular and widely used multimodal factorized high-order pooling mechanism to obtain a more distinguishable feature representation. Extensive experiments on three multimodal sentiment datasets (CMU-MOSEI, CMU-MOSI, and IEMOCAP) demonstrate that our approach reaches the state-of-the-art in an unaligned version setting. Compared with the mainstream methods, our proposed method shows superiority in both word-aligned and unaligned settings.

Aiming at the subtle differences between fabric defects and complex texture background and the problem of small size defects, we propose a fabric surface defect recognition method based on multibranch residual network. First, the convolution kernel is decomposed and the multibranch shared feature extraction module is constructed using the residual connection to mine the feature information at different scales. At the same time, squeeze-and-excitation network (SE-Net) with a reduction ratio of 16 is added to the module to suppress the interference of complex texture background. Second, the channel attention mechanism is embedded in the residual block of the network backbone to realize the compensation for the lost feature information. Finally, the swish activation function is used to enhance the accuracy and robustness of the deep network, and transfer learning is used to further improve the network accuracy. The experimental results show that the proposed model is superior to the existing model in terms of average recognition accuracy. The proposed model has the highest recognition accuracy of 96.17% for eight defect categories, which proves the effectiveness of fabric surface defect class recognition.

The adversarial examples have been proven to reveal the vulnerability of the deep neural networks (DNNs) model, which can be used to evaluate the performance and further improve the robustness of the model. Because text data is discrete, it is more difficult to generate adversarial examples in the natural language processing (NLP) domain than in the image domain. One of the challenges is that the generated adversarial text examples should maintain the correctness of grammar and the semantic similarity compared with the original texts. In this paper, we propose an adversarial text generation model, which generates high-quality adversarial text examples through an end-to-end model. Moreover, the adversarial text examples generated by our proposed model are embedded with watermarks, which can mark and trace the source of the generated adversarial text examples and prevent the model from being maliciously or illegally used. The experimental results show that the attack success rates of the proposed model can still reach higher than 88% even on the AG’s News dataset where generating adversarial text examples is more difficult. And the quality of adversarial text examples generated by the proposed model is higher than that of the baseline models. At the same time, because of the generated adversarial text examples are embedded with strong robust watermarks, the model can be better protected.

Fog generally causes loss or blurring of image information, significantly affecting work and life. This study proposes a three-stage fusion method comprising multiscale dark channel image fusion based on the dark channel prior, multiexposure image fusion using a gamma-correction technique, and a final fusion combining the restoration-based and enhancement-based methods to ensure good saturation and contrast in the dehazing result. The proposed method is experimentally evaluated in detail, and qualitative as well as quantitative analyses are performed. The obtained results indicate that our method can effectively improve the haze removal performance and is advantageous for dealing with long-distance hazy images.

With the enhancement of technology, data security has become incredibly significant. Nowadays, images and vital information are used in real-time, and high throughput architecture is required for internet of things and real-time applications. We mainly focus on the high throughput architecture of PRINT cipher. PRINT cipher is a lightweight block cipher with a block size of 48-bit and 96-bit. Loop-unrolled, 48-bit pipelined, and 96-bit pipelined architectures are designed and implemented on different field programmable gate array and application-specific integrated circuit platforms. The maximum operating frequency obtained for 48-bit pipelined is 259.67 MHz on Virtex-4 and 310.67 MHz for 96-bit pipelined on Virtex-5. Finally, with the help of a controller, 48-bit high throughput pipelined architecture was utilized for image encryption of both grayscale and color images. The security analysis of encrypted images for color and grayscale images using PRINT cipher shows better performance and more robust protection against statistical, entropy, and differential attacks.

Siamese-based trackers have achieved remarkable performance and widespread applications in visual tracking. However, most of the existing Siamese trackers are usually restricted by the difficulty of handling the misalignment between target localization and state estimation. To address the misalignment issue, we propose an efficient Siamese distribution-aware (DA) anchor-free (SiamDA) network for object tracking. First, the bounding box of regression targets is modeled by a flexible general distribution, avoiding the ambiguity of traditional rectangle representation in some complex scenes, such as occlusion and background clutter. On this basis, to eliminate the misalignment issue, a DA module is designed to connect the target localization and state estimation. In addition, the Pseudo-Intersection over Union sample assignment rule provides our anchor-free tracker with high-quality samples effectively during the training process, making it possible to further improve the tracking performance. Extensive experiments on official visual tracking benchmarks demonstrate the effectiveness and efficiency of SiamDA.

Image composition constructs a new image by cutting out a part of one image and then pasting it on another image. However, the quality of the new image is generally low due to inconsistency between the two parts. To overcome this, image harmonization aims to obtain realistic composite images by adjusting the color, illumination, and visual style of the foreground, to make it compatible with the background. Nevertheless, previous image harmonization techniques mostly concentrate on learning a mapping network from composite image to real image, and they ignore the significant role of background visual styles as a mapping guide. In this task, we consider image harmonization as a style transfer problem. Specifically, we take foreground as the content image and background as the style image, to transform the foreground’s style through background’s style features. To do so, we propose a unique self-attention-based module to learn the mapping between foreground features and background features using a modified self-attention mechanism. The proposed module can calculate the degree of correlation between the foreground and background according to the semantic distribution and also aligns the channel-wise mean and variance of the foreground features with that of the background features in the meantime. To comprehensively investigate the effectiveness of proposed module, we perform multiple experiments and ablation studies on the existing benchmark dataset iHarmony4. The experimental results prove that our module is more effective compared to the baseline in our task, and our harmonized image looks much closer to the real image.

Underwater robotics is an important part of oceanography, resource exploration, and marine engineering. The images captured by underwater robots are affected by environmental factors, such as scattering and absorption, which lead to color cast and haze issues. We propose a light-weight fusion-based convolution neural network (FCNN) that improves the visual content of the underwater image. In FCNN, two methods, such as automatic white balancing and contrast limited adaptive histogram equalization, are used in the preprocessing phase. Then the outputs of the preprocessing phase are fed into the multilayer neural network to learn the end-to-end features. Finally, the outputs of the preprocessing phase and multilayer neural network are fused to obtain enhanced image. The performance of FCNN is evaluated based on qualitative, quantitative analysis, and in terms of time complexity. The experimental analysis proves that the FCNN overcomes the existing conventional methods and deep learning-based networks.

Research into red, green, blue plus depth salient object detection (SOD) has identified the challenging problem of how to exploit raw depth features and fuse cross-modal (CM) information. To solve this problem, we propose an interactive nonlocal joint learning (INL-JL) network for quality RGB-D SOD. INL-JL benefits from three key components. First, we carry out joint learning to extract common features from RGB and depth images. Second, we adopt simple yet effective CM fusion blocks in lower levels while leveraging the proposed INL blocks in higher levels, aiming to purify the depth features and to make CM fusion more efficient. Third, we utilize a dense multiscale transfer strategy to infer saliency maps. INL-JL advances the state-of-the-art methods on five public datasets, demonstrating its power to promote the quality of RGB-D SOD.

RGB-T object tracking is a branch of visual tracking that has been widely applied to many fields, such as intelligent transportation and urban monitoring. Due to the interference of background clutter and occlusion, the existing trackers still suffer from the problems of unreasonable modal fusion strategy, insufficient feature extraction, and loss of semantic information. To solve these problems, we propose a residual learning-based two stream network for RGB-T object tracking. The overall feature extraction network is composed of three branches, and multi-layer convolutions are utilized to extract the features of visible, thermal, and fused images, respectively. First, aiming at improving the effectiveness of feature extraction, a weight generation module for hierarchical feature weight calculation is designed to guide the direction of feature fusion. Then, the residual block is employed to replace the single-layer convolution in order to increase the depth of the network, by which deeper semantic features are learned and the loss of semantic information is alleviated. Finally, a loss function with a penalty term is developed to adjust our network toward the direction of the best tracking performance of the dual modalities. This overcomes the negative impact of poor mode on model training. Experiments implemented on public RGB-T datasets indicate that our algorithm outperforms the recent state-of-the-art trackers in terms of both precision rate and success rate. Compared with the second-best comparison algorithm, our tracker improves the above two metrics by 0.4% and 1.5%, respectively. Our codes are available at https://github.com/MinjieWan/Residual-learning-based-two-stream-network-for-RGB-T-object-tracking.

Extracting vehicle information is highly significant to the development of intelligent transportation systems. Vehicle logo detection (VLD) plays an important role in the extraction of vehicle information. Traditional VLD methods rely on the detection of the license plate and have difficulty in accurately locating the logo. VLD methods based on deep learning have greatly improved detection accuracy and speed. To further increase the detection accuracy and speed of VLD, a VLD method, VLDetNet, is proposed. This method first introduces a combined multibranch block for VLD (denoted as vBlock), a new network block that can be combined with other network structures to greatly improve their recognition accuracy. By combining vBlock and MobileNetV1, we then construct a new backbone network structure, vMNet, which replaces the original backbone (DarkNet53) in YOLOv3. Finally, VLDetNet extracts spatial attention information from lower-level feature maps and uses it as spatial attention weights, to mine spatial relationships of pixels and facilitate vehicle logo extraction. To further promote research on VLD, we have labeled a new vehicle logo dataset, named HFUT-VL3, which contains 6000 images and 54 types of vehicle logos obtained from an real-time highway surveillance systems in China. We conducted experiments using both HFUT-VL3 and an open VLD-45 dataset. And the experimental results demonstrate that the proposed method achieves high detection accuracy and speed and outperforms most existing methods.

The discrete curvelet transform decomposes an image into a set of fundamental components that are distinguished by direction and size and a low-frequency representation. The curvelet representation of a natural image is approximately sparse; thus, it is useful for compressed sensing. However, with natural images, the low-frequency portion is seldom sparse. This manuscript presents a method to modify the redundant sparsifying transformation comprised of the wavelet and curvelet transforms to take advantage of this fact for compressed sensing image reconstruction. Instead of relying on sparsity for this low-frequency estimate, the Nyquist–Shannon sampling theorem specifies a rectangular region centered on the 0 frequency to be collected, which is used to generate a blurry estimate. A basis pursuit denoising problem is solved to determine the details with a modified sparsifying transformation. Improvements in quality are shown on magnetic resonance and optical images.

Graph convolutional neural networks (GCN) have been widely applied to many real-world problems and have achieved better results. To improve the accuracy of GCN in the three-dimensional human pose estimation (3D HPE) task, improved semantic GCNs (Simpre-RxSkip-NonCam-SemGCN and SPRA-SemGCN) are proposed. For the problem of insufficiently utilizing effective information in 3D HPE tasks, the dual-layer attention mechanism module (NonCam-attention) of extracting information is proposed. It establishes a spatial and channel-based fusion structure model by capturing the distance dependencies between different joint nodes of human poses and the associated features between different mapping channels. To reduce the training time and GPU memory for the preaggregated semantic graph convolutional networks (Pre-SemGCN) model, the simplified Pre-SemGCN module (Simpre-SemGCN) is suggested, which can accelerate the training speed of the model by eliminating the nonlinear activation layer. Further, a nested recursive expansion-based residual connection module (RxSkip + LN) is proposed to adjust the ratio of input and output with fewer parameters. To verify the effectiveness of the proposed model, we conduct experiments on the dataset Human 3.6M. The results show that the SPRA-SemGCN model proposed can effectively reduce the error of the 3D HPE task, and the mean per joint position error is finally reduced to 37.20 mm.

To better preserve the global semantic information and edge information of the original image during the multi-focus image fusion process, we propose an encoder–decoder network, which combines pixel loss function, multiscale structural similarity loss function and total variation loss function to further reduce the detail loss in the image reconstruction process. The model introduces dynamic convolution and global context network to improve its feature expression ability and global context modeling ability. In the fusion stage, the fusion strategy based on the edge feature map is used to fuse the two feature maps output by the encoder to obtain the final decision map. Combining the gradient information of the deep feature map and the edge feature map, the fusion strategy can more effectively extract the flat parts of the focused and defocused regions and enhance the edge features of the image. Finally, the proposed algorithm is compared with seven advanced fusion algorithms, and the results are superior to other fusion algorithms in subjective and objective evaluation indexes.

Deep neural networks (DNNs) have been shown to be susceptible to slightly perturbed adversarial examples. Even adversarial examples that are imperceptible to the human eye can easily lead to misclassification of DNNs. The most effective defense against adversarial examples is adversarial training (AT). Through AT, the robustness of the model against adversarial examples can be greatly improved. However, AT causes a decrease in the accuracy of the model’s classification of natural examples and serious overfitting problems. To improve the generalization ability of the model, we propose a method for enhancing model robustness through augmented data importance perception (ADIP). By extracting offensive adversarial examples for data augmentation and considering the attack strength of new adversarial examples to design an importance measure term to improve the loss function, the results of experiments show that the model robustness enhancement method based on ADIP improves the robustness of the model and alleviates the overfitting problem of the model caused by AT. In addition, our algorithm trains a robust model with a small computational cost.

Foreign object debris (FOD) on airport runways has always been a major problem in maintaining airport security. Currently, there are two main challenges in FOD detection based on computer vision. First, FOD is small and inconspicuous, which makes it difficult for the detector to find and categorize the object correctly. Second, airport pavements at night present a low-light situation, which makes FOD detection more difficult. To solve these problems, we propose an improved real-time detector bidirectional YOLO (Bi-YOLO), constructed by first adding a bidirectional PANet to YOLOv5. This adds a BiFPN-like weighted bidirectional operation to the PANet, allowing the network to improve the focus on small objects adaptively. Then, we use the anchor-free manner, SimOTA label assignment, and multibranch head to reduce the complexity of our model and improve the model’s performance in detecting small objects. Finally, we use special data augmentation strategies during training to improve the model’s performance on small objects in low-light situations. Experiments on the public dataset FOD-A and our own dataset FODInSues show that, compared with the popular object detectors YOLOv5, YOLOv3, CenterNet, and EfficientDet-D4, Bi-YOLO achieves the best performance in FOD detection, especially for small FOD in low-light situations.

Sketch-based image retrieval (SBIR) is a cross-domain retrieval problem whose main challenge is that sketches and images differ in many aspects, such as visual content and abstract degree. Existing approaches solve cross-domain gaps using deep neural networks (DNNs) to learn a joint global feature representation for sketches and images. To improve the retrieval effect using local features, some researchers have introduced attention mechanisms and multi-scale inputs. However, these cannot effectively locate regions of detail difference between sketch–image pairs, and the discriminative ability of these models is insufficient. Just as human observational habits range from global to local, this paper constructs a multiscale network with the help of attention modules. The network is stacked in a recurrent way and focuses stepwise on the most discriminative regions from coarse to fine through region-specific attention models. The part-aware attention module is specifically designed for each scale network to make it robust against local spatial misalignment. To address the challenges of SBIR, we extend the network into a three-channel structure that integrates the sketch, edge map, and RGB photo branches to acquire a more discriminative cross-domain feature representation. The proposed network is alternately optimized by weighted regularization triplet loss, classification loss, and pairwise ranking loss. We conduct extensive experiments on several commonly used SBIR datasets, quantitative, and qualitative results demonstrate that our network learns more fine-grained features and outperforms all baselines.

We proposed an innovative end-to-end generative adversarial network structure (CRVC-GAN) for video colorization, which is dedicated to solving the temporal consistency and coloring quality problems in the video coloring domain. For the temporal consistency problem, we utilize the network structure of recurrent neural network (RNN) because the model shape of RNN does not change with the input length and the calculation takes into account the history information to ensure the degree of correlation between successive frames. This better ensures the continuity between frames, so the consecutive frames get good temporal consistency in coloring. For the coloring quality problem, we utilize the UNet++ coding method to decode the video frames, integrating different levels of features to improve accuracy, and then enhance the object boundary information in the video frames, thus improving the coloring quality of grayscale video frames. In addition, we innovatively utilize the style transformation loss function in terms of images for model training to ensure temporal consistency and coloring quality by exploiting the slight difference in style variation from frame to frame. The experimental results show that, compared with the existing methods and through its own designed balance index, the colorization and temporary consistency balance index, CRVC-GAN obtained certain advantages on the DAVIS and Videvo public datasets with the results of 0.56176 and 0.74286, respectively, and achieved a better balance between time consistency and color quality.

The images captured under low-light conditions are characterized by low brightness and poor contrast, which affects the accuracy of computer vision tasks. In recent years, there have been a variety of low-light image enhancement (LLIE) models based on deep learning, but they have not been able to fully extract the multiscale information of multiple stages, resulting in poor generalization performance and instability of the model. Currently, a large number of multistage networks cause color distortion and stylization of images due to excessive transmission of noncritical information. To address these issues, we propose an LLIE via multistage feature fusion network. Our network consists of three stages. In the first two stages of the LLIE, S-UNet, which combines UNet and spatial weighted residual channel attention block (SWRCAB), helps the network extract more critical multiscale information and occupy a small amount of computing resources. In the third stage, we fuse the SWRCAB and a nonlocal sparse block into the original enhancement network to enhance the original resolution pixel-by-pixel. We also propose a fusion attention mechanism, which can provide real and effective supervision and control the transmission of a small amount of critical feature information for each stage. In addition, we add illumination guidance for image segmentation at the beginning of each stage of the network, excepting that the model can better focus on the dark part of the low-light image and avoid overexposure. We conduct experiments on multiple benchmark datasets to qualitatively and quantitatively demonstrate that the proposed method is more competitive than the state-of-the-art methods.

With the development of technology, it is becoming easier and easier for people to generate tampered images that are indistinguishable from the human eye, and the malicious use of these photos in news, academic papers, and criminal crimes has brought great harm to society. This paper proposes a deep convolutional neural network-based image forgery localization method to uncover the subtle differences between doctored images and real images. Specifically, the network achieves tampering localization through a three-stage enhancement scheme. First, the dilated convolution in the deep layer of the network is used to keep the feature map resolution constant, and the number of shallow convolutions is decreased to reduce the perceptual field, so the network focuses on local regions. Second, the feature enhancement module is used to fuse shallow features with deep features to effectively filter content information and highlight tampering features, making full use of local and global information to improve the generalization ability. Finally, the attention enhancement module reweights the convolutional feature maps in terms of channels and locations, respectively, to highlight the information regions around the forgery boundaries, thus guiding the network to capture more intrinsic features for image forgery. Extensive experimental results on several public datasets show that this method outperforms other state-of-the-art methods in image forgery localization.

With the development of cloud computing and the demand for privacy protection, reversible data hiding in encrypted images (RDHEI), which is a technology that embeds secret data in an encrypted image and can ensure error-free extraction of embedded data and lossless recovery of original images, has received widespread attention in recent years. In this paper, an high-capacity and secure RDHEI method is proposed. First, the pixel value of the original image is predicted. Using the properties of the Laplace distribution, the prediction errors are divided into two types of labels. Adaptive Huffman-encoded labeling is performed on these two labels, respectively. Pixel rearrangement is then used to reserve space where additional data can be embedded. Due to the reversibility of Huffman encoding and decoding, the receiver can achieve separable secret information extraction and lossless recovery of the original image. Experimental results show that compared with several existing methods, our method can achieve higher embedding capacity and good security, not only the invisibility of the image content, but also the undetectability of the secret information.

Skin cancer is the world’s fifth most prevalent cancer. Skin cancer is seen in one out of every three cancers. The quality of early and precise skin cancer detection is quite difficult. Some studies have been conducted on the computerized detection of malignancy in skin lesion images. The manual tuning of network architecture design is heavily reliant on domain expertise and is tough. In the present research, a one-shot neural architecture search technique is proposed that boost the search efficiency by modeling NAS as a one-shot training process of an over-parameterized supernet. As a result, various architectures can be derived from the supernet, and share the same weights. An evolutionary memetic algorithm employing local search was proposed to optimum the network architecture, which yields minimal computation time and accurate classification of benign and malignant skin lesions. Our search algorithm comprises all conceivable networks in each solution space. This approach prevents the neural architecture from being scratched. The proposed work proved efficient in ResNet search space with better outcomes than manually created networks on the ISBI dataset, which resulted in an accuracy of 93.26% and the computation cost is decreased. The search process also lasts an average of 18 h on a graphical processing unit.

Visible-infrared (IR) person re-identification is a technology that matches the identity of the same person in two modalities. The main challenge is to discover the differentiations between different identities and the similarities between the two modalities. To solve this problem, we propose a cross-modality consistency learning network, which jointly considers cross-modal learning and distillation learning. It consists of two associated components: the feature adaptation network (FANet) and the modality learning module (MLM). The FANet combines global and local information to extract more discriminative features on the same identity images, and MLM is used to alleviate modal differences between visible and IR images. Our model could adaptively select the high-quality person image according to the potential contribution of each image, to avoid negative knowledge transfer. Extensive experiments on the public SYSU-MM01 and RegDB datasets demonstrate the superiority of our approach over the current state-of-the-art technologies.

Spatial color algorithms (SCAs) are computer vision procedures widely used for image enhancement and human vision modeling. The main characteristic of SCA family is that they mimic the behavior of the human vision system (HVS), achieving in this way robustness and the capability to adjust their effect according to the image content. Here, we review 35 different, popular Retinex-inspired SCAs discussing and providing a set of measures for their evaluation in terms of image quality. To this purpose, we also introduce SCA-30, a real-world color image dataset made publicly available. The algorithms considered here include and spread from well-known Retinex implementations, Retinex variants, Milano–Retinex and related inspired enhancers, illumination/decomposition approaches, and deep learning-based techniques. Data and code used for the evaluation are made freely available to the community, to pursue further analysis and comparisons.

Surface defects in hot-rolled steel strips are one of the common product problems for the steel industry, which harm the product appearance, affect the corrosion and wear resistance, and shorten the product service life. The natural defect samples are sparse, category imbalanced, and expensive manual annotations. Therefore, it is crucial to study the data augmentation and classification methods for small sample surface defects. To solve the above problems and improve the accuracy and real-time performance of defect classification, we propose a random offline data augmentation algorithm (Random-CutMix) and an improved MobileNet architecture (SP-MobileNet). The Random-CutMix algorithm expands the dataset by random sampling to balance the number of each defect class. The SP-MobileNet combines the inverse residual module with the channel shuffle mechanism (CSIn-Module) and pyramid split attention (PSA) module, which facilitates cross-group information flow and improves model representation capability and generalization performance with low computational cost. The accuracy, recall, F1 score, parameter, computational complexity, and frame rate of SP-MobileNet with Random-CutMix on the X-SDD dataset were 95.97%, 95.22%, 95.46%, 6.5 M, and 0.54 G, 72 FPS, respectively. The experiment results indicate that our method outperforms the state-of-the-art methods and provides an effective trade-off between accuracy and instantaneity in actual industrial production.

Flame detection is critical to human society. Current flame detection methods have problems about huge model volume and large parameters. We proposed an improved YOLOv5-light that reduces the parameters while maintaining the same detection accuracy. First, we use the channel separation shuffling method to ensure the information exchange between channels and optimize the residual structure. Second, depthwise separable convolution is used as the downsampling module in the backbone network to reduce the model’s parameters further. Finally, the Ghost module is used to improve the feature extraction module in the head network to reduce computational cost caused by ordinary convolution. The final experimental results show that the proposed YOLOv5-light network achieves detection accuracy similar to that of YOLOv5, with a 65.6% decrease in parameters and a model volume of only 34.8% of the latter.

RGB-D salient object detection (SOD) can usually be divided into three stages: feature extraction, feature fusion, and feature prediction. Most approaches treat the feature information extracted by the backbone network identically in the final two stages of detection, neglecting the fact that various modalities and different hierarchical features play distinct roles in SOD, resulting in poor detection results. To solve this problem, we propose a transformer-based difference fusion network (TDF-Net) for RGB-D SOD that treats modal features and hierarchical features differently in the feature fusion and feature prediction stages, respectively. First, we adopt the pyramid vision transformer as a feature extractor to obtain hierarchical features from the input RGB images and depth images, respectively. Second, we propose a differential interactive fusion module, in which the RGB modality and the depth modality learn modality-specific features independently, and the two modalities guide each other to fuse features. Finally, we divide the hierarchical features after cross-modal fusion into high-level and low-level features and propose three types of cross-layer fusion modules to discriminately integrate features from different layers to predict the salient maps. Extensive experiments on five benchmark datasets confirm that our proposed TDF-Net outperforms the state-of-the-art methods.

The number of Arabic handwriting documents has increased greatly. Studies conducted in the Arabic handwriting recognition field have progressed significantly in recent years in different areas. The existing studies for Arabic language compared to others such as Latin remains insufficient. During the last decade, neural networks (NNs) have become the de facto standard for deep learning, and the combination of two or more NNs has proved its ability to learn complex objects, such as handwriting, and emulate human brains. We proposed an approach combining convolutional NN and bidirectional long short-term memory for recognition. This unique approach makes it possible to recognize Arabic handwriting words without segmentation. The proposed architecture is very efficient in terms of accuracy for Arabic word recognition. The hyperparameters set for our model were chosen based on a structured process of randomness and grid search to increase the accuracy of the model.

In recent years, segmentation-based two-stage neural networks have emerged in the defect detection domain. Although these neural networks have shown great success in surface defect detection, they suffer from blurred edges, noise, and imbalanced data, which degrade the detection performance. In this study, we address these limitations by presenting a new two-stage defect detection framework. To solve the problem of defect detection with blurred edges, we propose leveraging multilevel representations during segmentation. Furthermore, we employ a joint loss function that combines the binary cross-entropy loss and Dice loss to reduce the influence of noise. Meanwhile, the joint loss function that contributes to our model is more robust for imbalanced data and is better for segmenting small-sized defects. To verify the performance of our proposed model, comprehensive experiments were conducted on two surface defect datasets: the Kolektor surface-defect dataset (KSDD) ( http://www.vicos.si/Downloads/KolektorSDD) and the Deutsche Arbeitsgemeinschaft für Mustererkennung dataset (DAGM) ( https://hci.iwr.uni-heidelberg.de/content/weakly-supervised-learning-industrial-optical-inspection). By performing more accurate segmentation of defects, our model shows superior results over the compared networks; it provided a 100% average precision on KSDD and 100% mean accuracy on DAGM, which demonstrates the effectiveness of our model.

Modeling the visual features of the forearm internal rotation angle during Tai Chi training is difficult, and its accurate detection is lacking. To address this issue, this paper proposes an interactive visualization analysis method based on Harris corner detection and edge contour feature extraction to visualize and adjust the internal rotation of the forearm during Tai Chi training in real time. In the proposed method, the visual images of limb movement and forearm rotation during the Tai Chi training are collected as features, and the wavelet image denoising method is used to denoise the images to improve the signal-to-noise ratio of the output. Meanwhile, the image pixel uniform traversal method is used to continuously traverse the image subblocks to extract the edge contour features during the forearm internal rotation interaction. By means of superpixel feature decomposition and grayscale information superposition, the fitting effect of the forearm internal rotation angle is fed back, and based on this, the forearm internal rotation angle is modified to realize the interactive visual analysis of the forearm internal rotation angle. Experimental results show that, compared with existing methods, the interactive visualization performance of the proposed method is better. The proposed method can fit, track, and solve the rotation angle in real time; the required analysis training images are more accurate, and the Tai Chi training data detection and dynamic analysis abilities are improved.

The aim of image fusion is to merge image information from multiple source images into one image. The fused image can contain more complete and clear scene information without introducing artifacts. The results of image fusion are required to be faithful to the original image contents. However, we notice that the unfaithful fusions often contain hallucinated content, which cannot be aligned with any input images. For the first time, we propose a full-scale connected-based fusion network (FSCF-Net) for infrared and visible image fusion to achieve the goal. To make full use of the multiscale features of the image, the full-scale skip connection is used in the decoder network of FSCF-NET to fuse the full-scale deep features extracted by the encoder. To enforce faithfulness, the FSCF-Net adopts a hybrid loss function consisting of detail-preserving loss and structure-preserving loss. To effectively fuse the features of different source images, a spatial-channel attention-based feature fusion strategy (SCAF) is proposed. The SCAF fusion strategy measures the importance of features from both spatial and channel dimensions. The detailed analysis is provided about the proposed model in our experiments that are performed on publicly available datasets. We adopt subjective evaluation and 6 objective metrics to compare the proposed FSCF-Net with 10 state-of-the-art fusion algorithms. The experiments demonstrate that our method achieves the best comprehensive performance on the overall metrics.

Erratum corrects an error in an author’s name.