This PDF file contains the front matter associated with SPIE Proceedings Volume 12287, including the Title Page, Copyright information, Table of Contents, and Committee Page.

The object detection algorithm has experienced the transition from the traditional algorithm to the object detection method based on convolutional neural network. The traditional object detection algorithm mainly includes VJ detector, HOG, DPM, etc. It mainly uses the sliding window method to detect and locate the object, and there is a certain bottleneck in the running speed. With the increase of training data and computer performance, the object detection algorithm based on deep learning has a good effect in both detection accuracy and detection speed, and has become the mainstream algorithm of object detection. Through the author's own understanding, this paper summarizes various object detection algorithms in a more understandable language.

Since 2019, the COVID-19 has been hanging over the whole world, causing uncountable financial loss. In this regard, wearing masks becomes a precaution for the public. However, some people are wearing masks in a wrong way, which may cause virus infection. To detect the wrong wearing of masks, we use 3 classic Convolutional Neural Networks, namely LeNet-5, AlexNet, and VGGNet-16, based on a unique dataset, to train the model and analyze the results. On the unique dataset, LeNet-5 achieved an accuracy of 80.3%, which was the lowest among the three networks, AlexNet attained an accuracy of 90.6%, which is near the precision of VGGNet-16, 92.83%. This work may help the advance of a digital city, making COVID-19 precaution under control.

The research is mainly about the Python implementation of three basic algorithms: Depth-First Search, Breadth-First Search and A* Search. Besides, the paper introduces some of the auxiliary knowledge and theories: the four criteria for measuring algorithms, three kinds of queues, and grid-based simulation. While other coding languages like C++ and Java may focus on algorithm more and may have lots of templates and other solved works, Python seem not to focus so much on this, and OOP (object-oriented programming) is the major topic of high-level Python coding. Thus, constructing algorithms and mapping (or simulating real-world situations) indeed become a challenge.In this work, two simulations are made by generating classes: 1. grid-based world 2. graph and vertex. Grid based world introduces coordinates and heuristic functions, which constructs the environment suitable for greedy algorithms. Graph and vertex use nodes to represent different states in the of the world, and using frontier-based search, the nodes will be explored step by step from the initial state until the goal state is reached. After the environment and algorithms are constructed, finally, analysis for the three algorithms are presented in the conclusion part according to the criteria mentioned.

With the rapid development of Internet technology, network threats are also increasing rapidly. Traditional ways of dealing with network threats, such as intrusion detection technology and malicious code detection technology, reveal the shortcomings of poor quantitative analysis and poor prediction of future situations. In order to cope with today's complex and changing network situation, network security situation awareness technology has gained more attention. Network security situation awareness consists of three layers: situation element extraction, situation understanding and situation prediction, and a complete network security situation awareness model is constructed through the three parts. The paper introduces the concept and development of network security situation awareness, describes the current mainstream algorithms and their advantages and disadvantages for each layer, and finally summarizes and outlooks on the problems of current mainstream algorithms and the trends of future development.

As the carrier of electronic components, PCB plays a role in supporting interconnection in electronic equipment. With the continuous in-depth development of the PCB market, PCB defects have become more and more subtle, complex and diverse. The traditional quality inspection operations are complex, lack for intelligence and have low precision, which greatly reduces the inspection efficiency. In this paper, a target detection method based on YOLOV3 network is designed for six defects of PCB board: open circuit, short circuit, excess copper, and etching gap. Through the design of the convolution block and the extraction and design of the feature layer, the backbone model of the darknet53 unique to YOLOV3 is generated and trained in three batches. And use the non-maximum suppression method to obtain the position, score and type of the defect frame, so as to achieve defect detection, and on the basis of only 2000 training data, the overall detection success rate is 79.03%.

For noisy indoor house type images obtained by infrared scanning and other methods, they usually need to be further optimized before they can be used. This paper proposes an improved RANSAC algorithm based on the conventional RANSAC algorithm,and combines the least squares method to detect,extract and enhance lines of noisy indoor house type image.The binarization of the image is completed by threshold judgment division to avoid the excessive dispersion of pixels, which leads to the deviation of the subsequent RANSAC extraction model.Then, a recursive-based desalination method is proposed to desalinate the image, desalt the noise and improve the accuracy.When detecting a straight line, a sequential recording method is proposed to record and process the multi-segment line segments that are in the same straight line but not connected to ensure that there will be no omission or error. Finally, the extraction of multiple straight lines and line segments is realized by cyclically removing interior points, so as to complete the straight line detection and extraction enhancement process of the entire noisy indoor house type image. The actual results show that the method has high accuracy and timeliness, and can make the line detection and extraction completion rate reach about 90%, and can still maintain a very low error rate even in a heavily noisy environment.

The purpose of this paper is to discuss some 5G technologies and to emphasize the nature of uav-assisted 5g networks by extending the value of FBS deployment. Research to address network-related issues such as energy constraints, optimal 3D deployment, and path planning will be discussed. Finally, this paper will study FBS trajectory optimization and MBS optimal cell edge layout.

As a new type of advanced simulation teaching aids and interactive methods, the virtual simulation system attracts the love of relevant demanders and uses them for learning and research with its unique novel and convenient advantages. In this paper, an experimental system of practical computer virtual simulation technology is designed. The experimental area is developed in C# language using Java tools, and is programmed with design ideas such as object-oriented modeling, support for interactivity and automatic control functions. In terms of software, SQL statements are used to build a model to complete the whole training process to simulate the real world; in terms of content, Web pages and database are combined to build a knowledge network learning environment (BPO); a simple online test system is provided through the course Q&A platform for users to do use and to give corresponding answers to different versions of questions.

NFV systems are becoming more complicated with various of software modules, vendors and interfaces introduced. Testing such systems is very time consuming and requires lots of expertise. In this paper, we propose an efficient automatic testing framework to simplify and speedup the testing process. There are two critical components in this framework, i.e., a structured and reusable test case format and an abstraction layer to adapt different testing tools. Based on this framework, a tester can quickly compose the test cases by just setting the corresponding parameters of the test tools and target components. After this setting, the framework will automatically arrange the testing order and invoke the testing tools to fulfil the test. The experiment on real systems shows that our framework can greatly improve the testing efficiency.

The advent of sign language has given the deaf a more accurate and convenient way to express themselves, but not everyone knows sign language. Hearing-impaired and deaf people still find it difficult to communicate with normal people without interpreters. Therefore, the application of computer vision technology to America Sign Language letter (ASL) recognition will greatly facilitate such silent communication. However, the development of ASL recognition programs requires engineers to have extensive knowledge of sign language, which is very difficult to achieve. To solve this limitation, our study applied automatic machine learning (AutoML) technology to ASL recognition based on the Google Cloud Platform (GCP). Firstly, we standardized the labels and augmented some of the images. Subsequently, the preprocessed data set was used to train the motion detection model. During training the model, we built and normalized the model to improve the accuracy of the model. Finally, the accuracy and recall rate of the proposed method achieved 99.5% and 92.7%, respectively. In addition, the model was also deployed to an Android system for testing.

Recently, federated learning has received a lot of attention. It is a novel distributed machine learning scheme that enables a large number of edge computing devices to collaboratively learn and train models without leaking local data. Federated learning can break down data barriers, effectively solve the problem of data island, and is committed to ensuring user privacy and data security, and has received extensive attention from industry and academia. However, in practical industrial scenarios, federated learning needs to consider more effects of heterogeneity than traditional machine learning. Aiming at the problems of high communication cost and heterogeneous client data in federated learning, this paper proposes a federated learning training method based on knowledge distillation, which realizes the sample-level information exchange between the server and the client, and guides each client to train local personality. The model is optimized to reduce the impact of heterogeneous data on the client side, and the distillation loss is used to ensure the stability of the model performance. Theoretical analysis and experimental results show that the federated learning algorithm proposed in this paper can ensure the data privacy and security of participants. And in the scenario of Non-IID data, the algorithm can improve the training efficiency and reduce the system communication cost on the premise of ensuring the accuracy of the model, and has the feasibility of practical industrial application.

Machine learning (ML) has been widely used in various scenes, such as face recognition, automatic driving and image recognition and other fields. However, in these applications, attackers usually exploit various vulnerabilities in machine learning models to mine private information, including model training information and private data information. The attacks against machine learning training data mainly include membership inference attack (MIA). The attacker of membership inference attack will judge whether the data sample is in the training data set of the target model given the data sample and the target model. Most of the existing member inference attacks are based on confidence attacks. Researchers found that if the confidence returned is hidden, they can effectively defend against these attacks. Therefore, the latest research proposed label-only attack, but this attack requires hundreds of queries for a sample, and abnormal query accesses are easy to be found by the target model. In this paper, we propose a member inference attack based on the transferability of adversarial samples. The adversarial samples generated under different models of the same task have certain mobility characteristics, so as to use the shadow model to generate adversarial samples for attack and reduce the attack cost. The evaluation shows that under the restriction of fewer queries, only the label output will still reveal private information, and the performance of the member inference attack based on the transferability of adversarial samples is better than that based on confidence in most cases.

The identification of pig number is very important in large-scale breeding. The traditional manual identification is time-consuming and laborious, and more difficult when the natural environment and pig behavior change. In order to solve this problem, a pig identification method based on 5G and convolutional neural network was proposed. Firstly, 5g network is used to collect and transmit the image data, then the enhanced prewitt Algorithm based on FPGA chip is used to detect the edge, finally, 5000 images containing pig and 5000 images without pig are used to detect the edge, training and testing the built convolutional neural network. The experiment shows that the method can recognize the pig outline under different angles, illumination and posture with an accuracy of 94.3% , which provides technical support for intelligent weight measurement, estrous behavior detection and early diagnosis and warning of epidemic diseases, improving the economic efficiency and management efficiency of pig-raising.

The International Classification of Diseases (ICD) is a health care Classification system initiated by the World Health Organization. ICD codes are used to quantify important statistical data and facilitate the search for patient cohort with similar diagnosis. In addition, they are also of great value and significance as a means of standardized information exchange between hospitals. Manual ICD coding is a time-consuming and laborious work, now most people use machine/deep learning methods for automatic coding. TextCNN, TextRNN and TextRCNN have been the mainstream models of multilabel text classification task since they were proposed. In this paper, three language models are combined with ICD automatic coding task. The experimental results show that the three models have achieved good results. In addition to that, this paper demonstrates the limiting factors of the model performance through detailed experiments to guide the future work to make further progress.

When making network security situation prediction, researchers usually use historical network security situation values to make predictions. Such predictive models often do not perform well on complex networks with multi-source inputs. Aiming at the problem that the prediction accuracy of the existing prediction models is not high in complex networks, a network security situation prediction model based on grey wolf optimization algorithm to optimize support vector machine is proposed. In this paper, the situation elements are first accumulated, and then the support vector machine is used to independently predict different situation elements in the network, and finally the predicted values are fused into the network security situation value. The parameters of the support vector machine are determined using the grey wolf optimization algorithm. The simulation results show that the model has a good prediction effect.

Information retrieval is the foundation of content-driven applications, and the quality of the search results directly affects whether users can quickly obtain the required information. At present, vertical search engines for specific fields meet the needs of users to obtain specific information to some extents, but full-text search engines based on text matching cannot search from the perspective of semantics, so that search results strongly dependent on the selection of search terms. Aiming at the above problems, a scoring and sorting algorithm based on search term expansion is proposed. This method combines the research contents of search term expansion based on relational ontology and semantic similarity calculation based on word vector, discusses the search strategy for learning data library, and proposes a reasonable expansion method for search terms in vertical search engines. The Dscore ranking algorithm of search results proposes a new optimization strategy for the ranking algorithm of search results, realizing a search engine more suitable for library applications.

Aiming at the problem of efficient allocation of radio resources for massive users in Non-Orthogonal Multiple Access (NOMA) systems, a user-subchannel bilateral dynamic matching algorithm is proposed. First, a joint user-subchannel and power allocation optimization problem for multi-carrier NOMA systems is established with the target of maximizing system energy efficiency. The user-subchannel is modeled as a bilateral many-to-many matching game, the GaleShapley matching algorithm is introduced, and an efficient two-sided matching algorithm is proposed to achieve the dynamic and stable matching of the user-subchannel. Then, on this basis, the power allocation problem is allocated to the power resources through the convex difference (DC) planning algorithm. From the simulation results. Compared with the traditional matching algorithm, equal power allocation algorithm and orthogonal frequency division multiple access scheme, the proposed algorithm can improve the system energy efficiency and converge faster.

It is important to observe whether people wear masks in public due to the spread of Convid-19 in the world. In this work, I built the object detection model to classify whether people are wearing facial masks and wearing them correctly with Automated Machine Learning (AutoML) technology on easyDL platform. The whole process is completed automatically and it is easy to operate so that one does not have to be an expert of machine learning to train a qualified model for practical use. The dataset includes images and labels in XML format. I adjusted a few parameters such as the data-augmentation operator and the algorithm provided by the platform to choose. Finally, an acceptable result was achieved and I deployed the model on the web with a web framework called Flask. The model offers a relatively satisfactory result and can be successfully called by API provided by the platform.

With the continuous growth of private transportation, parking has become a great difficulty in densely populated areas. This paper is committed to realizing vehicle autonomous parking under controlled and limited external complex conditions. Considering the continuity of state and action space, TD3 and DDPG algorithms based on actor critical architecture are adopted. The paper uses two algorithms to train the agent in simulated parking environment. The simulation results show that the DDPG algorithm based on DRL can achieve high-precision parking path planning.

To research the problems of the rolling bearing fault diagnosis under strong noise, an enhanced convolution gated recurrent neural network was proposed, which is mainly composed of convolution neural network and gated recurrent neural network. In the proposed method, a convolution neural network is used to extract relevant features from vibration signals, and a gated recurrent neural network is used to further process relevant features to realize the diagnosis of bearing fault and its severity in complex scenes. In order to further enhance the ability of the network to deal with noise, a signal input method based on random sampling strategy is proposed, and the activation function in convolution network is improved, so as to enhance the bearing fault diagnosis ability of the network in complex scenes. The experiment on the Case Western Reserve University public data set proves that the proposed network framework could achieve the leading bearing fault diagnosis accuracy in complex scenarios such as high noise.

Aiming at the problem that the cluster head selection of clustering routing protocol in wireless sensor networks is unreasonable, resulting in uneven network load and shortened life cycle, a cluster head selection algorithm based on Levy flight Butterfly Optimization Algorithm is proposed. Firstly, based on the Butterfly Optimization Algorithm, the Levy flight operator is introduced to improve the search ability of the Butterfly Optimization Algorithm. In the process of cluster head selection, a new fitness function is designed according to the two factors of node residual energy and distance between nodes. The improved Butterfly Optimization Algorithm is used to improve the random problem of cluster head selection and comprehensively select the better cluster head. Our method considers the factors such as node energy and distance comprehensively. In contrast, the simulation results prove to have better effect in the lifetime and energy consumption.

In this research, we find out current traffic road sign detection is mostly based on high precision detection device to collect road data for machine learning to train model for autonomous driving. For most people, createML makes it easier to use and create models such as our road sign detection than using scientific equipment in the lab. If more and more people have access to machine learning, then it will cost us less to develop a model. The point of doing this research is to provide a solution to reduce the cost of future road sign detection and development models that are valuable for real-life applications, such as Carplay. We are using an easier approach from using current exist road sign pictures, and use create ML to do the model training and find out if could be potential solutions for transplanting models into phone apps to connect car device for autonomous driving.

This paper proposes an optimization method for distributed energy access to distribution network based on chaotic genetic simulated annealing algorithm, and establishes a joint optimization with the total network loss of the power grid system, the minimum node voltage, the average voltage deviation, and the user power purchase cost as the objective function. model, to realize the location selection and constant volume of distributed energy, and to use the weighted method to realize the simplification of the objective function. The chaotic genetic simulated annealing algorithm is used to solve the optimization model, and the simulation comparison is made based on the IEEE-69 node system and the particle swarm algorithm , the results show that the proposed method has higher effectiveness and reliability.

Unsupervised domain adaptation (UDA) person re-identification (re-ID) aims to apply useful knowledge learned on labeled source domain datasets to unlabeled target domain datasets. Most successful UDA re-ID methods combine clustering to generate pseudo labels for feature learning and further fine-tuning on the target domain in an alternating manner. However, the interaction of the two steps is offline, which may make the noisy pseudo labels greatly hinder the classification and retrieval ability of the whole model. In order to purify these noisy pseudo labels, in this paper, a framework called Unsupervised Confident Co-promoting (UCC) is proposed. Specially, two peer teach-student co-training models are adopted simultaneously to online refine the noisy pseudo labels with the offline clustering algorithm and supervise each other during iterations. More significantly, we introduce a confidence strategy that greatly improves the confidence of generating pseudo labels in the case of multi-network collaboratively guided learning. The combination of the above two allows our final method to greatly improve noisy pseudo labels on re-ID task, achieve a huge performance boost and generalize to more deep learning domains. Moreover, our method achieves a significant improvement in common evaluation indicators in the four most common re-ID experiments compared to the state-of-the-art (SOTA) methods, and even on some results is comparable to the supervised learning method.

With the continuous development of cloud computing technology, the application of cloud computing technology in all walks of life has become a trend. Combined with the current situation of university computer laboratory construction, this paper introduces the application scheme of cloud computing technology in computer laboratory construction and management from the perspective of cloud computing, virtualization and other technologies. Compared with the traditional laboratory construction scheme, the laboratory system based on cloud computing can effectively reduce the laboratory construction and management cost, improve the management efficiency and resource utilization, and achieve the sharing of experimental resources. Students can experiment on demand anytime and anywhere.

For enterprises, their applications have different types and complex configuration files. Carrying a large number of applications that depend on different environments on a server cluster will bring huge difficulties to the deployment and maintenance process. In recent years, with the development of virtual machines and container virtualization technologies, more and more enterprises have used IaaS cloud platforms or container cloud platforms to carry their applications. Compared with the IaaS cloud platform, the container cloud platform has higher resource utilization and stronger observability, but there are challenges with high barriers for getting started. In order to overcome this challenge, this paper designs and implements a lightweight application management architecture based on container cloud. This architecture is based on various open source tools and self-developed modules to achieve "lightweight" operation and maintenance of cloud applications.

Malaria is a transmittable disease caused by the parasites which belong to the Plasmodium family which is spread by the bite of the female mosquito. In the past, they were detected by trained microscopists who analyze microscopic blood smear images. However in recent years, a series of deep learning methods had been appeared. In this passage, we use ResNet with different parameters to classify Malaria cell images. We preprocess a dataset with more than 27,000 Malaria cell images from more than 350 patients. On this dataset, we train a convolutional neural network which outputs the classification accuracy and loss. By using transfer learning, we find the best network architecture in which we can classify the cells with the accuracy of 94.93%.

In order to improve the use efficiency of university laboratory information resources, strengthen the smooth progress of teaching and research tasks and ensure the reliability of private cloud, this paper proposes a deployment and monitoring scheme of private cloud system with highly reliable software and hardware resources, in which the monitoring scheme is highlighted. Adopt virtualization and other key technologies to ensure the reliability of physical host and virtualized environment. The deployment of a small private cloud platform composed of 6 hosts and equipped with high-capacity storage devices is realized, which can meet the daily teaching and research management needs of teachers and students in the laboratory. Taking ZABBIX as the framework, expand and improve its monitoring function, complete the fault detection and analysis of physical equipment, virtual resources and software applications of private cloud platform, focus on the monitoring of the hardware foundation layer, and realize the reliability monitoring of private cloud platform. The scheme takes the self-built private cloud of small laboratories in colleges and universities as the research subject and the open-source monitoring software as the means to realize the effective utilization and reliability assurance of resources. While meeting the requirements of daily teaching and research tasks of laboratories, the scheme is also applicable to the management and monitoring of self-built cloud of small and medium-sized enterprises.

Language, the most apparent distinction between the human and the animal, has been focused by the data scientists in recent decades. They are dying to discover a method that the human terms can be translated into machine language. After the seething of the English recognition, more researchers focus on the Chinese recognition both in China and abroad. However, differing with the English recognition, Chinese recognition is more challenging because of the tremendous character categories. Moreover, compared with single character, Chinese text recognition can be more challenging due to the join-up and the different writing styles. In this article, three algorithms about the Chinese text recognition are going to be illustrated and evaluated, including the segmentation recognition and the no segmentation. In the segmentation recognition part, we will describe briefly about the single character recognition because the segmentation recognition process includes this representative method, and we also introduce a classical algorithm about Chinese text recognition written by Wang in 2012 [1]. In the no segmentation part, we display two kinds of ways to solve the problem. One is based on HMM and Another is based on FCRN.

Aiming at the problems of low-illuminance image brightness, low contrast, and lack of detailed information, a multi-scale weighted image enhancement algorithm based on HSV space is proposed. The algorithm first converts the image from the RGB space to the HSV space; then uses the information that can characterize the image characteristics in the H and S channels of the HSV space as weights to enhance the V channel; finally, the guided filtering is used to enhance the details of the image to obtain a strong contrast, Images with sharp edges and details. To verify the superiority of the algorithm, a comparative experiment was carried out, and the enhancement algorithm was objectively compared through three indicators of En (Entropy), EME (Enhancement by Entropy), and AG (Average Gradient). The test results show that the indicators of low-illuminance images have been significantly enhanced, from the original 6.32, 15.53, and 2.87 to 8.35, 75.89, and 6.58 respectively, and compared with the existing methods, the contrast and detail information of the restored results Verifies the effectiveness of the proposed method.

In recent years, object detection, as an important branch of computer vision, has made great progress and received attention from various fields. However, for remote sensing images with small size, arbitrary orientation and dense distribution characteristics, there are still big challenges. In this paper, we propose a multi-classification detection network for small and dense remote sensing small targets. A coarse-grained target cluster anchor frame is designed to delineate the range of target regions on remote sensing images, which effectively reduces the influence of the large-scale background on the detection network for the subsequent detection network. Meanwhile, according to the concept of multi-scale detection in SSD network, a dense convolution is applied to the multi-scale feature map after the backbone network to improve the features of small targets. Finally, the rotating anchor frame is used as the precise location of this detector, and the Focal LOSS function is used to mitigate the gradient disappearance problem in the regression process. Different panels in this model are extensively experimented on the remote sensing dataset DOTA and NWPU VHR-10 to get more excellent detection results.

Scientific measurement of taxi passenger volume is of great significance for optimizing urban traffic. In this paper, based on the satellite positioning data of taxis in Huangshan City, the identification method of taxi passenger trips is studied with Python as the analysis tool, and the taxi passenger trip trajectory is extracted. Through the sampling survey, the passenger load factor of different periods is calculated, based on which the distribution characteristics of taxis under different time and space is analyzed and visualized through ArcGIS and TransCAD. The research demonstrates that the average daily passenger volume of taxis in Huangshan City is 14,013 person-times, and the number of taxi passengers on rest days is higher than that on working days, and significantly higher during the day than at night; taxi trips are mainly distributed in the main urban areas of the city, accounting for 87.36% of the total trips, forming two passenger corridors, east-west and north-south.

At present, blockchain users lack effective means of supervision and authorization for intermediaries at different levels, making user data privacy at risk of leakage. Aiming at the above problems, this paper studies the trusted supervision technology scheme suitable for blockchain data privacy protection, enriches and expands on the classic BIP32 protocol, and designs and implements a dynamic authorization strategy in the new protocol. Authorization policies make private data truly in the hands of users by changing and setting regulatory permissions. The experimental results and performance comparison show that the blockchain data privacy protection method proposed by this scheme is efficient, stable, and has strong practicability, which is suitable for data privacy protection scenarios of trusted supervision. At the same time, experiments show that with the increase of the encryption key depth, the supervision time increases linearly.

Assessment of which place was damaged after a hurricane is an emergent task. However, handling this task by humans is laborious and inefficient even with the help of hurricane satellite imagery. A good way to detect damaged buildings effectively and time-savingly is to apply a pre-trained convolutional neural network (CNN) based on post-hurricane satellite images for processing real-time satellite images. However, training a complex CNN is time-consuming, and choosing different networks could get different efficiencies in training the network and detecting damaged buildings. In this paper, we compare the efficiency of different CNN architectures that VGG16, VGG19, ResNet50, InceptionV3, and convolutional neural network built by us on binarily classifying damaged buildings on post-hurricane satellite images. We evaluate the efficiency of networks by the balanced and unbalanced tests accuracy, the training time, and the inference time. With the given data for training and testing, the network we built is significantly more efficient than other existing networks with less training time, less evaluation time, and more accuracy (over 99%); but it is less complex and has lower parameters compared to other existing networks.

Cloud-native technologies have received extensive attention from operators, and some network functions have been deployed in containers. The network function is usually responsible for the function of data forwarding, so the data plane of CNF is a very important technology, which has attracted the common attention of the academic and industrial circles. At present, there are many researches on traditional virtual machine-based network functions, but less research on containerized network functions, especially on the data plane performance of CNF. In this paper, we focus on the data plane performance of CNF. On the one hand, we compared the impact of data plane acceleration software on CNF performance through experiments. The conclusion shows that data plane acceleration software has greatly improved the bandwidth and packet forwarding capability of CNF. On the other hand, we also tested the impact of different data plane acceleration drivers. We found that different DPDK data plane acceleration drivers have very little impact on CNF performance, and only have a certain impact on small packets (<256 Byte).

In view of the singleness of traditional text emotion analysis, multimodal emotion analysis extends emotion analysis to the analysis level of text, image, sound and so on. In order to use the context interaction information expressed in each modality, a multimodal interactive emotion classification model based on video context is proposed in this paper. The ALBERTBiGRU network structure is built for text feature learning, and the independent BIGRU model is used to extract context features from text, audio and video modality. Based on the attention mechanism, the emotion analysis task is completed after the fusion of the three modal features. Compared with the existing models on MOSI and IEMOCAP data sets, the accuracy and F1 value of emotion classification reached 81.71% and 81.44 on MOSI data set, 66.97% and 67.20 on IEMOCAP data set, which were 1.41% and 2.16% higher than the highest benchmark value respectively, effectively improving the accuracy of multimodal emotion prediction.

Intelligent warehouses greatly improve management efficiency and reduce the management cost of enterprises. Now many enterprises are building their own intelligent warehouses. However, the current intelligent warehouse lacks the communication between digital space and physical space, and the efficiency of warehouse management still has a lot of room for improvement. The digital twin technology is a useful way to realize the fusion of virtual warehouse and physical warehouse and optimizes the activities. In order to realize the communication between digital space and physical space of the intelligent warehouse, eliminate information island and improve the efficiency of warehouse management, this paper proposed the digital twin system architecture of intelligent warehouse based on digital twin technology. At the same time, this paper analyzed the system application requirements, designed and constructed an intelligent warehouse digital twin system. Finally, the prototype system is developed with the warehouse of a garment enterprise as an example to verify the feasibility and effectiveness of the system.

While targeting at the classification of images of un-/damaged post-hurricane buildings, this paper investigates into the effects of different data augmentation techniques on three popular convolution neural network models, together with a model built by our team. The main derived results are the test accuracies for each combination of data augmentations and models, showing the competency of a certain combination on our task; thus, comparisons and evaluations are made. The conclusion drawn is that, firstly, convolution neural network models are capable of completing the classification with a rather high rate of success. And it is shown that there could possibly exist such data augmentation techniques that do not improve the performance of a model.

Classification of damaged buildings post hurricanes could be a cinch with the implementation of deep learning networks. Many researchers have engaged in conducting such tasks with the help of neural networks on satellite images of posthurricane buildings, while few of them grant appropriate attention to the data preprocessing methods even considering some unique properties of the satellite images. In this paper, we investigate several traditional data preprocessing approaches and some novel image preprocessing methods while making comparisons and analyses. Eventually, we could reach an accuracy over 98% with appropriate preprocessing approaches with a simplified ResNet network. Furthermore, we also analyze the effectiveness of the data augmentation method for damaged building classification, including geometric transformation, photometric transformation, and image mixing.

When hurricane happens, immediate and accurate assessment on housing damage is critical for rescuers. It is essential for the government to detect and measure the number and rating of the damaged buildings. Traditional way is to process this task manually, which is time-consuming. Recently, with the achievement made in the field of deep learning, this task could be done more easily using Convolutional Neural Networks. In this paper, we aim to classify the damaged buildings post hurricane using CNN models. We experiment with four different models and propose the final model-- SECure. SECure utilizes SE block to generate channel attention for feature representations, leading to a better performance. SECure achieves 97.95% accuracy on validation set, which can show a significant performance on handling this task. We hope our innovation can be facilitate the usage of CNN in real-world rescue scenarios.

Combine federated learning with the mobile network a good opportunity for the introduction of AI technology research and application in the mobile communication network.This article proposes a kind of the mobile network architecture based on FL to promote the integration of FL technologies with mobile communication network and to meet the joint modelling needs of users, operators, and industries. And points out some key technologies of the mobile network to illustrate how the mobile network to support FL working. Meanwhile, proposes some use cases to illustrate what possible application scenarios and values, if the FL technology is introduced into the mobile communication network.

In recent years, with the continuous progress of the power transmission project, the scale of the transmission line network in the region continues to expand, the difficulty of the inspection of the transmission channel in the region gradually increases. In order to improve the inspection accuracy and efficiency of the transmission channel, this paper uses the BJ-2 high-resolution satellite image data as the data source, and adopts multi-scale segmentation technology. Technology and object-oriented classification algorithm, study and analyze the extraction method of key ground feature information of transmission channel, the research shows that: (1) object-oriented algorithm is suitable for the information extraction of high-resolution image data, its ground feature information extraction accuracy is high, it can meet the inspection work of transmission channel, but it still needs the intervention of manual interpretation work, which shows that transmission based on high-resolution satellite image technology channel inspection is feasible; (2) the accuracy of shed information extracted by multi-scale segmentation technology coupled with object-oriented classification algorithm can reach 95%, which can replace manual interpretation work, while the accuracy of road information can be relatively low, but it can still replace most of manual interpretation work; compared with traditional field inspection and manual visual interpretation, multi-scale segmentation technology coupled with object-oriented classification algorithm has the advantages of high boundary precision and high efficiency.

Hyperspectral images (HSIs) classification methods based on convolutional neural networks are becoming increasingly popular. Many proposed methods extract spatial and spectral features simultaneously, and the interaction between the two types of features leads to unsatisfied classification results. Moreover, most of the existing CNN-based methods mainly consider single-scale, which can cause some important information to be neglected. To address the aforementioned two issues, we propose a Attention-Based Multi-Scale Network (AMSN) for HSIs classification. First, the proposed network is based on 3D-CNN, through channel branch and spatial branch, the AMSN can capture more distinctive spatial and spectral features using different convolution kernels, respectively. Second, Local and global features are extracted by dense network. Then, two features are concatenated to make full use of multi-scale features. Third, attention blocks are used after each branch to achieve the most distinctive features. Experimental results on three HSIs datasets demonstrate that the proposed framework can achieve better classification performance than the several state-of-the-art methods.

In complex traffic scenarios, MEC (Mobile Edge Computing) servers need to compute and allocate autonomous driving tasks and communication resources in an efficient and timely manner. Under complex traffic flow, there are problems such as repeated computation, redundancy, and waste of sensory information required by vehicles with similar driving tasks when a single vehicle is used as the service object. To address this problem, our research constructs a task-driven MEC system that encodes driving task features based on position, velocity, and acceleration using the spatial association hypothesis and proposes an improved clustering grouping algorithm based on the task similarity of nodes in a traffic flow. To save the computational and communication resources of the server, we further propose a core vehicle election strategy for the optimal channel state and use spatial coordinate transformation to realize the sensory data sharing between the core vehicle and subsequent vehicles. The MEC server only needs to maintain simple data frame communication with all vehicle nodes to realize real-time sensing of the traffic flow topology. The task-driven edge computing system built in our study fully utilizes the computational power of the vehicle terminals themselves, the channel resources between vehicles, and the autonomous sensing capability of the vehicle sensors. Our research provides a solution to the problem of resource exhaustion and communication delays in high-density traffic flow scenarios, such as accident-induced congestion and complex intersections.

With the development of social platforms, people are more likely to use pictures to share their sentiments and opinions. With the development of deep learning, the current mainstream image sentiment analysis research mainly focuses on the use of high-level semantic sentiment features of images, and less on the use of low-level sentiment features such as colors, lines and so on. Aiming at the above problems, this paper proposes an image sentiment analysis model based on multi-level feature fusion (MLFF). Firstly, the convolution neural network is used to learn different levels of image features; Then the multi-level features are input into the bidirectional sequential neural network, and attention is introduced to fuse to generate multi-level fusion features; Finally, the multi-level fusion features are input to the full connection layer for classification. The experimental results show that the fusion model proposed in this paper can make effective use of multi-level image features and effectively improve the performance of image sentiment analysis.

Previous works have shown that the effectiveness and reliability of CNN used in computer vision is closely dependent on whether the dataset is adequate. In this paper, we illustrated how data augmentation, a technique which expands original datasets through transformations, could help improve CNN models’ performance and prevent overfitting through several experiments. After importing two extremely small datasets found on Kaggle, we carried out experiments by comparing the performance of models employing and not employing data augmentation techniques, while keeping track of their training accuracy, validation accuracy and their final test accuracy respectively. Eventually, noticeable improvement could be witnessed when employing appropriate data augmentation techniques on small datasets in our experiments, as such methods could generate a more comprehensive and sufficient dataset to help models learn the key features.

Graph convolutional networks (GCNs) have achieved great success in hyperspectral image (HSI) classification. However, there are also some difficulties that need to be solved, such as the lack of enough labeled samples in the training process and the difference between the numbers of different land-cover analogies is too large, which will lead to poor classification performance. In order to alleviate those problems, we propose a Mutual Learning Graph Convolutional Network (called MLGCN) with an imbalance loss, which trains two GCN simultaneously, and these two GCNs can learn from each other by using mutual learning strategy in the training process. Experiments on three real data sets show that the proposed MLGCN method achieves state-of-the-art results in terms of three classification evaluation metrics, including overall accuracy (OA), average accuracy (AA), kappa coefficient (κ),.

The main improvement of image steganography is that it can generate dense images with high visual quality. The generated countermeasure network based on least squares is more stable than the conventional generated countermeasure network and converges faster than WGAN. The main work of the least squares generation countermeasure network is to replace the cross-entropy loss function with the least squares loss function. This is considered to improve two problems of the traditional Gan, that is, the image quality generated by the traditional Gan is not high, and the training process is very unstable. The least squares generation countermeasure network attempts to use different distance metrics to build a more stable, convergent and high-quality countermeasure network. In the process of network training, the human visual model is used to evaluate and restrict the visual quality of the secret image, and enhance the concealment of steganography from the perspective of improving the visual quality of the secret image. In the process of steganalysis of the secret image, the secret image is evaluated in combination with the method of image forensics, so as to ensure the visual quality of the image and reduce the trace of image modification, Enhance the concealment of image steganography.

In the path planning problem based on geographic information system, distance and resource overhead often conflict with each other. For example, the shortest path may not be the most time-saving or convenient one. So, a Multi-objective vehicle path planning method is needed. Therefore, this paper proposes a vehicle path planning algorithm based on improved deep reinforcement learning, which not only solves the problem of single optimization objective restriction, but also reduce the chance of falling into local optimum in traditional path planning algorithm. In this paper, DQN network structure is selected, and neural network is used to store action value function instead of Q table. In the training process, prioritized experience replay strategy is used to accelerate the convergence. The simulation experiment and analysis of the proposed model show that the improved DQN algorithm model can effectively find the path with the optimal value, making the path planning more rational and intelligent.

The ergonomic design study of artificial intelligence lower limb-assisted brace for the elderly is a new design standard of lower limb-assisted brace for the elderly with mobility problems. Based on Human Factors Engineering, this study tested and analyzed the advantages and disadvantages of human lower limb motion mechanics, human gait motion law and existing lower limb assisted brace design cases at home and abroad, and concluded that the common external assisted method is less man-machine efficient than the internal assisted method. Therefore, a new brace joint rotation curvature, component parameters and other key information was designed based on the structure of the medial assistance method. Subsequently, Rhino software was used for digital modeling, physical prototyping, experimental testing and analysis of the design solution and continuous optimization of the design. At the same time, the perceptual engineering design method was utilized to meet the humanized aesthetic design requirements. The prototype of the design study was finally completed, which is more in line with the evaluation criteria of "human-machine-environment system" than the existing market design in terms of functional rationality, human-machine performance and human experience. This demonstrates the validity of the design method and is an important reference for the design standard of the lower limb support for the elderly.

In recent years, with the development of the Internet of things, the data collection methods are more and more abundant and the data structure is more complex. Anomaly detection by analyzing multi-source data has become a research hotspot. Multi source data is robust in data accuracy, but there are still some problems, such as difficulty in effective fusion and feature extraction. This paper proposes an anomaly detection method based on multi-source heterogeneous data fusion from the perspective of sequence. The extracted subsequences are mapped into the feature subspace, and then a unified fusion feature space matrix based on multi view is constructed. Finally, an anomaly scoring method based on property attribute fusion is proposed. Anomaly detection is carried out based on fuzzy clustering. The accuracy of the algorithm is verified by many experiments.

Image Classification is a fundamental task that attempts to comprehend an entire image as a whole. Using a suitable optimizer and useful learning rate adjustment strategy is very important in this task. Recently some papers have discussed the effects of different learning rate adjustment strategies combined with different neural networks. However, there is little discussion about the effects of learning rate adjustment strategies combined with different optimizers. In this paper, we compare the effects of different combinations of different optimizers and learning rate adjustment strategies on image classification task. After analyzing the experimental results, we found that the combination of RMSProp and ExponentialLR obtained the best result. The combination increases the training speed and achieves high accuracy.

Due to the rapid development of computer and communication technology, a large number of delay-sensitive and computation-intensive applications have appeared in the mobile Internet. Traditional cloud computing cannot handle such applications well. Mobile edge computing (MEC) based on distribution is considered to be an effective technology to solve this problem. However, the computation resources in edge base stations are limited, and computation offloading and resource allocation are the key factors affecting service efficiency and energy saving. In this paper, a Markov decision model is constructed for the adaptive computation offloading and resource allocation environment. Deep reinforcement learning (DRL) algorithms do not require prior knowledge, and can learn optimal strategies through training to solve dynamic decision-making problems, and the Deep Deterministic Policy Gradient (DDPG) algorithm based on DRL is used to solve the model to obtain computation offloading and resource allocation strategies that minimizes time delay and energy consumption. The simulation results show that, compared with other algorithms, the computational task completion rate of our proposed algorithm is greatly improved, and the time delay and energy consumption are effectively reduced.

In the aftermath of a hurricane, damage assessment is critical for relief helpers. One way to measure damage is to detect and quantify the number of damaged structures, usually by driving around the affected area and noting them down manually. This process can be labor-intensive and time-consuming and is not the most efficient method. Therefore, in this paper, we choose to apply image classification to improve the efficiency of assessing damaged building. From the public satellite imagery data of building, we extract square-sized images and create training, validation, test, and unbalanced test datasets. Each square-sized image contains a building to be classified as either ‘damage’ or ' no damage’. We use the existing convolutional neural network model to classify the datasets, and then use the results to modify our algorithm to achieve better programs. We apply our final model to the dataset that is satellite image from Texas after Hurricane Harvey with 99% accuracy to demonstrate the efficiency of assessing damaged buildings.

This paper proposes a method to calculate the distance between vehicles from the perspective of the car, in which monocular vision information from the driving recorders of the host car is used to calculate the distance between the host car and the car ahead. The author first detects all the vehicles with their type in an image input, sifts out the front vehicle from them, and locates the license plate from the bounding box of the vehicle. Then the paper designs an algorithm to measure the length of the license plate in the bounding box of the license plate and utilizes the common characteristic of license plates in America and in other countries that all the plates of one type of vehicles in one country share to derive a novel method to convert the length of the plate in the image into distance between the driving recorder and the license plate, therefore obtaining car-to-car distance.

With the gradual deepening of construction engineering informatization and the acceleration of office electronization in the architecture field, the data barrier will be progressively disintegrated. Regarding the research on the collaborative performance of large-scale BIM model under the new situation, we present the massive BIM sharing and interchanging distributed framework (MBSIDF). MBSIDF first uses the scientific system design concept to build a distributed basic service framework, then establishes a large data centralized sharing platform to increase the efficiency of data aggregation and interaction, and finally builds a massive BIM model real-time analysis platform to improve data processing performance. Experiments show that MBSIDF can build an elastic distributed BIM collaboration platform, realize the centralized management function of massive BIM models, and significantly improve the processing performance of massive BIM models.

Many platforms based on auto machine learning approaches have emerged as a result of the rapid development of machine learning. However, the performance differences between platforms have rarely been investigated. In this study, Google Cloud AutoML and Apple CreateML platforms are compared to find the difference in terms of the performance. To be more specific, we trained the model with the same dataset from Kaggle Animal 10. Three experiments are conducted between Google Cloud AutoML and Apple CreateML to achieve this goal. Finally, we arrived at the conclusion by examining the training accuracy, validation accuracy, and average precision across these two platforms. According to each pair of comparative tests, Google Cloud AutoML appears to perform better in every experiment when compared to Apple CreateML.

For disaster emergency response, higy-definition satellite imagery is quite important. However, what’s more important is machine learning and other advanced technologies which can use these images to get much more essential information. In this paper, we use transfer learning to classify buildings post hurricane. First, we propose a new model based on VGG-16, which can be more suitable to the experiment environment and get better performance. Second, we compare the efficiency and time the model use among different models in processing this task. We find that our model based on VGG-16 get the best accuracy which can reach to over 99%.

Accurate prediction of gas loads is of great significance to the rational planning of gas consumption by gas companies. Aiming at the problem that traditional forecasting methods cannot model the variation law of the load itself and its influencing factors at the same time, a load forecasting method based on long short-term memory (LSTM) is proposed. Firstly, the paper considers the historical load and various influencing factors comprehensively. Secondly, the temporal ability of LSTM is used to identify the internal variation on load in a longer time range. Then, the load prediction performance of different network architectures based on actual data is verified and compared with other algorithms. The results show that the proposed method can effectively improve the accuracy of load forecasting.

As global electricity consumption continues to grow, carbon-neutral strategies have been proposed by various countries. The integrated power system is an important tool and development direction to assess the power situation in the future. However, most countries today focus on the traditional single-indicator power system, failing to consider the coupling relationship between different indicators in the integrated power system. Considering the disadvantages of traditional big data downscaling with overcrowding among clusters, this paper will use the unified flow approximation and projection (UMAP) method, while considering the coupling relationship among multiple indicators, to build an atomic search optimization method based on chaos algorithm to optimize the kernel function, which can improve the prediction accuracy by more than 50% compared with the traditional BP neural network. Finally, a multivariate load short-term forecasting system for integrated power system management platform is built based on the current popular technical architecture, and a real-time forecasting system with front and back-end separation is constructed based on SpringCloud microservice.

Web services have been integrated into every aspect of people's lives, and both online shopping and entertainment have extremely high requirements for Web service quality, so it is crucial to design a prediction algorithm that meets users' quality Web service quality for user experience. Existing algorithms mainly rely on the contextual information of users and services, without using the hidden graph structure information between users and services, and without considering the association between service quality attribute values, resulting in poor prediction accuracy for sparse data. To address this situation, a model based on graph attention network to capture information between users and services is proposed, and combined with MMOE multi-task learning to simultaneously complete the prediction of two quality of service attribute values simultaneously.

The drug quality re-inspector is an application of machine vision artificial intelligence technology to drug quality reinspection. In this paper, the author proposes a lightweight deep convolutional neural network machine learning model using a custom-trained YOLO machine learning algorithm to detect the integrity of pharmaceuticals. More importantly, it uses artificial neural networks (ANNs) to model complex relationships in the data. A color drug image segmentation method using a combination of HSV model and color distance is used to address the case of uniformly colored drug attachment in drug images. After testing, the accuracy rate was significantly improved compared to the traditional manual rechecking, eventually reaching 86%.

It has been recognized that power control strategy of spectrum sharing for USVs(unmanned surface vessels) counts. The policy for power control of a major USV is proposed to increase profit of spectrum resources. In the environment of an USV sailing on the sea, it is essential for the major USV to communicate effectively with the minor USV. We propose a learning-based power control approach for the major USV to share the joint spectrum with the minor USV. Specifically, the major USV updates its transmitted power abide by a power control policy that is defined beforehand. The minor USV is unaware of the major USV’s condition of transmitting power. The proposed method exploiting the strengths of the Deep Q-Network (DQN), could help the minor USV adjust its transmitting power and make full use of the empty spectrum. Both can transmit information favourably accompanied by the satisfactory service quality. The results of our experiment show that the minor USV communicate with the major USV effectively through our method. Compared to the traditional reinforcement learning method DQN, our method achieves better results.

At present, the uneven distribution of entities and the low frequency of some entities in medical text data leads to the low accuracy of medical named entity recognition. To solve the above problems, a neural network model based on dictionary and mutual attention (DB-MA-BiLSTM-CRF) is proposed. The model includes Bert embedding layer, BiLSTM-CNN layer, mutual attention layer and CRF layer. The medical dictionary is fused in the Bert embedding layer to inject medical vocabulary information; Then it is input into the BiLSTM-CNN network layer to extract the global and local features of the text respectively; The features extracted from BiLSTM-CNN network layer are spliced into the mutual attention layer for weighted extraction of important features. Compared with the two benchmark models, the experimental results show that the model proposed in this paper has strong advantages.

With the advancement of information technology, deep learning and face recognition technology is widely employed in financial transactions, public security management, employee attendance, and public security. Regarding the technical advantages of the recognition system, this paper first explains the demand of the age for the system, then studies the impact of deep learning on face recognition, and finally highlights the application path and future development direction of the face recognition system in public security for the reference of relevant personnel. The face recognition system upgraded by deep learning has increased personal information security and favorable influences public security management.

Neural networks have been successfully applied in many fields. However, satisfactory results can only be obtained under large sample conditions. When it comes to small training sets, the performance may not be very good, or the learning task may not even be completed. In this study, our main task is to use data augmentation to solve the problems caused by small datasets. And we chose ImageDataGenerator in keras to obtain countless derived pictures for training. Also, we use the current state-of-the-art method, EfficientNet, to construct a classification model of different optimizers. The best model accuracy is 0.9647. Finally, we prove that EfficientNet is one of the most suitable methods for this study.

Air pollution is an environmental problem facing mankind today. Therefore, predicting the concentration of air pollutants in advance plays an important role in people's life and government decision-making. In this paper, a multi-channel asymmetric structure prediction model based on temporal convolutional network (TCN) is proposed. As TCN omits some feature information when learning time series features, increasing the number of channels will improve the receptive field of the model, cover longer historical information and extract more time series features. The influence of meteorological factors on the concentration of air pollutants is fully considered in the prediction model, which is used as an auxiliary factor to improve the prediction performance of the model. The concentration of air pollutants collected from the air monitoring station in Fushun City, Liaoning Province, is used as the data set to verify the effectiveness of the model, and the experimental comparison with other prediction models is conducted. The results show that the model proposed in this paper has more accurate prediction accuracy and stronger stability.

The fine-tune paradigm adopted by traditional paraphrase recognition tasks cannot fully exploit the knowledge of pretrained language models (PLMs). At this stage, the Prompt paradigm reconstructs downstream tasks by constructing templates to make it more suitable for the training form of PLM. However, there are many processes, the code base is not supervised, and a single discrete template limits the model prediction ability. In response to this problem, this paper proposes a method of introducing hybrid prompts based on OpenPrompt. OpenPrompt makes the Prompt process have a unified framework, and hybrid prompts solve the problem that discrete templates cannot fully mine PLM knowledge. This paper first constructs a hybrid template with [mask] slots, and then transforms the original input through the template to obtain x_prompt , and then x_prompt is input to the bert-base-chinese model of the multi-layer bidirectional transformer based on the attention mechanism for training. When the model trains the optimal prompt, calculate the label with the highest probability of filling in the label set, and finally map the label to the prediction result. The experimental results show that compared with the fine-tune paradigm, the F1 value of the prompt paradigm exceeds 4.69% under the same PLM and hyperparameters. Compared with not using soft prompt, the average accuracy and average F1 value obtained with soft hints are 2.294% and 2.31% higher, respectively. Meanwhile, when the number of soft hints is 6, the accuracy and F1 value reach the highest.

In this paper, we aim to predict the cases of covid-19 pandemic according to linear regression model and random forest model. We decide to try to predict the virus using the daily high and low temperatures because it is one of the biggest factors that can affect the spread and death of the virus.we decide to use days_num,vaccine_days,and ma_temp_high as features.Cases and deaths as labels. We find that that the virus surely has some relationship with temperature. If the theory turns out to be true, in the future, adjusting control efforts based on temperature could greatly improve efficiency and save money. Reduce ineffective expenditures and improve the quality of prevention and control.

Traffic forecasting plays an important role in intelligent traffic system. Forecasting traffic data in the future will provide great convenience in our daily life, such as avoiding congested roads in advance. In recent times, many methods for traffic prediction have been proposed, but most of these methods use complete data sets for prediction, and seldom pay attention to the sparse spatiotemporal data sets. Some recent studies mostly complete the data first before the prediction of sparse data. Therefore, this paper proposes Attention based Spatio-Temporal Generative Adversarial Network (ASTGAN) to solve this problem. ASTGAN uses the attention mechanism to preprocess the sparse data to utilize the temporal dependency to pre-complete the data, and then input the processed data into an mask graph convolutional recurrent network, which further complete the data with its spatial correlations and provide forecasting results. In order to ensure the accuracy and authenticity of the prediction, we also use generative adversarial network. Experiments on real data demonstrate the effectiveness of our method.

As a new treatment for malignant tumors, highly specific targeted drugs have been vigorously promoted in modern medical cancer treatment. At present, the problems of large single dose, high quality standard and high cost of targeted drugs have become well-known rich drugs. Accurate IC₅₀ prediction of targeted drugs has important practical significance for cancer economy and precision treatment. In this paper, 179 groups of characteristic parameters of targeted drugs against hepatoma cell line from Genomics of Drug Sensitivity in Cancer (GDSC) open source database are selected as model input. Half maximal inhibitor concentration (IC₅₀) of targeted drug action is used as model output. Then, IC50 prediction model based on SSA-BP neural network is established. Compared with other methods, the proposed model has faster convergence speed and higher prediction accuracy, and the overall regression coefficient is 0.96463.

During the COVID-19 pandemic, wearing gauze masks was proven to prevent people from infection. In public areas like shopping malls or schools need a way to supervise people wearing masks. This research aims to provide managers of public areas with an idea to solve this problem by GoogLeNet which is a type of convolutional neural network algorithm. Especially in crowded public areas, people should wear masks whether for their health or the health of others. These areas, such as stations and shopping malls, can only supervise people wearing masks at the entrance, but it is difficult to supervise people wearing masks inside buildings. As a result, many people will take off their masks or incorrectly wear them indoors due to heat. In this case, we consider how to intercept everyone's avatars in the video on closed-circuit television. Use neural network training algorithms to monitor everyone's mask-wearing situation. And promptly warn people who wear masks incorrectly or who do not wear masks.

Aiming at the problems of low accuracy and long detection time of traditional fire detection algorithms, a fire detection method DAI-YOLO based on improved YOLOv3 is proposed. On the basis of the YOLOv3 algorithm, the network structure is improved, the scale of the input image is increased, the traditional convolution is replaced with a depthwise separable convolution structure, the model parameters are reduced, and the detection rate is improved; the multi-scale feature detection is used to increase the shallow detection scale and add 4 times upsampling feature fusion structure to improve fire detection accuracy; optimize k-means clustering algorithm. The experimental results show that the accuracy and recall rate of the improved algorithm are 91.2% and 84.2%, the mAP is as high as 84.6%, and the detection speed can reach 0.31 s, which can meet the real-time and accuracy of fire detection.

Rice is one of the most important crops in the world and the most important food crop in most areas. It is very susceptible to pests during the growing process, which affects the yield and quality of the rice. Accurately assessing the types of rice pests is the basis for pest eradication. A method for detecting rice pests based on improved Varifocal Net is proposed. A self-attention mechanism is introduced to improve and refine feature maps and a group normalization method is added. The experimental results show that the method can achieve accurate detection of 9 rice pests such as rice leaf roller and rice borer, with a mean average precision of 96.3%, which is 1.9% higher than the original Varifocal Net.

In this paper, we use the image data in the dataset-FER2013 as a sample. First, we format the CSV file in the dataset. After finding that there is a large difference in the number of different types of pictures, in order to compare and verify the impact of the number difference on the results, we enhanced the data set through the proposed data augmentation method, so that the data set can meet the experimental needs. After that, by understanding the artificial intelligence learning algorithms at home and abroad, we decided to establish a training model using AlexNet network, and divided the pictures to be identified into three groups with different numbers: original data set, down sampling data set and dataset with data augmentation. In the experiment, we achieves 52.3% accuracy after data enhancement, and 66.9% accuracy using the down sampling dataset.

In this paper, we present an end-to-end model based on modified Bidirectional Encoder Representations from Transformers (BERT) for Chinese named entity recognition (NER) in natural language processing. The model is composed of the SpanBERT layer and the Conditional Random Field (CRF) layer. By using combination, the model can express the input characters in the better form of "word embeddings", eliminating the steps of feature engineering or data processing in conventional approaches, and can be widely applied to the task of Chinese NER. Our experiments demonstrate that the SpanBERT-CRF model can effectively utilize the contextual data features and give more accurate recognition results. On our data set, the SpanBERT-CRF model had excellent performance with a recognition accuracy of 91.33%, outperforming the benchmark NER model BiLSTM-CRF (Bidirectional Long Short Term Memory, Conditional Random Field) and BERT-CRF model in performance and F1 score.

With the wide application of network data in many fields, network representation learning technology has become the focus of scholars' research. Especially in dynamic networks, how to combine network structure information with time information is the difficulty of current research. This paper proposes a dynamic network representation learning method (ESNA) based on temporal neighborhood aggregation. Further, We divide the influence of nodes on the network into local power and global power, learn the local influence of nodes on the network by using time difference attention mechanism aggregation in the temporal neighborhood, delimit the influence range of nodes on the global network by using random walk model outside the time series neighborhood, and finally fuse the local and global influence of nodes on the network based on word vector model to form a network embedded with time information. We experimented with real dynamic network data in the link prediction task, and the results showed the effectiveness of our method.

Sentiment recognition is an important part of sentiment computing. To address the problem that traditional Bert models are studied only in textual aspects and a single model has significant uncertainty, an emotion classification recognition method using CM-BERT and Bi-LSTM-DenseNet dual-model decision fusion is proposed. Using both feature-level fusion and decision-level fusion strategies, the superiority of the dual model in multimodal sentiment recognition is highlighted, and the conclusion that the dual model outperforms the single model in multimodal recognition is verified. The model implemented in this paper achieves 64.5% accuracy on a homemade multimodal dataset, a 1% improvement compared to the CM-BERT model, and a 7% improvement compared to the Bi-LSTM-DenseNet model.

In the current situation of a severe epidemic, the feasibility of model training on the automated machine learning platform and deployment on mobile terminal for mask detection is feasible since it may provide higher accuracy. In this study, we optimized the original dataset, adjusted the quantity proportion of various types, screened the image quality to alleviate the effect of the overfitting. The dataset is made into an open source that supports different formats and has been uploaded to Kaggle. Then, the trained models based on Create ML and EasyDL platforms are deployed in the mobile terminal, the accuracy of each model and detection effect on mobile terminals were obtained, respectively. Lastly, the advantages and disadvantages of the models and applications of the two platforms (Create ML and EasyDL) are compared.

Rapid and automatic extraction of buildings from remote sensing images can provide an auxiliary decision-making basis for urban management, military investigation, and post-disaster emergency assessment. Based on the deep learning model UNet++, this paper integrates the EfficientNet and SGE attention, combined with the auxiliary training of crossentropy loss function, so that the model can segment the buildings in remote sensing images. The experimental results on the Inria Aerial Image Labeling dataset show that the evaluation indexes of the proposed model are high in five cities, with the highest OA of 96.83%, mIoU of 0.80, and Kappa of 0.76. The buildings in the segmentation results are complete and accurate, and the boundary is clear.

Emotion plays a very important role in People's Daily life: a wealth of emotions is a great help to the speaker in conveying ideas.At the same time, video emotion recognition is one of the important research fields of human-computer interaction.In recent years, a lot of work on large-scale video emotion recognition has promoted the rapid development of image emotion recognition methods and systems.Based on this, this paper introduces the video emotion description model, summarizes the video description is often used in the process of emotion recognition characteristics, use at the time of the video feature extraction, processing method and the classification of the need when video emotion recognition algorithm, and has carried on the comparative analysis, hoping to promote the development of emotion recognition technology.

Short term load forecasting is an important guarantee to ensure the smooth operation of power system, but there are many factors affecting it, which makes it difficult to predict it whit high-precision in the future. In order to improve the accuracy of load forecasting, a day ahead load forecasting model based on Spearman correlation coefficient, MultiTaskLasso and Multilayer Perceptron (MLP) is proposed in this paper. Firstly, Spearman correlation coefficient is used to select the leading steps in order to eliminate the low influencing factors related to load; Secondly, MultiTaskLasso is applied for feature extraction to extract the optimal feature set effectively that can best express the load fluctuation from different influencing factors with different candidate input variables; Finally, MLP is applied to predict the load with 48 lagging steps in the next 24 hours. Model comparison and verification results show that, the predict accuracy has increased by at least 24.38% and 25.32% respectively when the selection of the leading steps and the extraction of the optimal feature set have completed, and the prediction accuracy of Spearman-MultiTaskLasso-MLP is improved by at least 39.42%. The effect of the leading steps selection and feature extraction is obvious, and the prediction accuracy of each comparative model is improved significantly. The results of this study can provide a reference for the selection of leading steps and feature extraction for similar studies.

Recently, COVID-19 has become a world-wide pandemic with high infection and death rate. Researches mainly focus on how to predict the trend or control the infection. In this paper, we aim to predict COVID-19 pandemic using physical signs. We apply three methods, including decision tree, random forest and supported vector machine. We claim that all the methods achieves a satisfying performance, with the top accuracy is 99.8%. Our method provides a significant contribution in prediction COVID-19 infection, which will apply in real-world application with a real-time inference time.

The research of link prediction is widely used in many fields such as social relations and biological sciences. So it has become a very important research direction. Existing link prediction methods are usually based on the assumption that "the more similar two nodes are, the higher the probability that a link exists between them". That is, the similarity between nodes is considered as the existence probability of a link between nodes. Such an approach does not fully exploit the hidden information in the network, so the accuracy of prediction needs to be further improved. In this paper, we propose a link prediction model based on collaborative filtering. The model combines the similarities between nodes and the known network topology by the method of collaborative filtering, which effectively improves the accuracy of link prediction. Extensive experiments on six real-world datasets demonstrate the validity of the proposed model.

To protect the environment, governments introduced many policies about garbage recycling and garbage classification. However, the different garbage classification policies between different areas bring many inconveniences for people who come from other areas. In the meantime, the development of garbage classification application requires developers to have lots of experience about object classification. In this work, we employed Automated Machine Learning for the garbage classification application, which is called AGC method. The AGC method uses Google AutoML service to train the garbage classification model with Baidu AI studio garbage dataset. During train the model, we set the data augmentation and normalization to increase the precision of the model. To compare model effects, the model trained by AutoML was compared with MobileNet. At the end, the model was deployed on an iPhone to test the model. Our method achieves 0.92 in precision, which is much higher than MobileNet, and the application correctly recognized the real-word garbage image with 0.87 confidence.

With the development of information technology and intelligent manufacturing technology, more and more industrial control system (ICS) devices with IP address can be accessed through the Internet. In order to quickly search and monitor the networked industrial control device or system, and master the type, scale and regional distribution of the device, it is urgent to use relevant technologies to identify the networked industrial control system or device. In this paper, we first analyze the industrial control protocols widely used in industrial control systems and the special fields in the protocols. Then, we design a detection scheme to collect the data packets of ICS device, and select the specific word segment of industrial control protocol according to the diversity and consistency of field values. Finally, we propose an automatic detection system for industrial control device, which uses convolutional neural network (CNN) and principal component analysis (PCA) to extract feature field and construct feature fingerprint, and uses limit learning machine (ELM) to realize fine-grained ICS device recognition.

The huge pressure of market demand and competitive environment makes supply chain finance the choice of most enterprises. The emergence of public health emergencies such as the COVID-19 epidemic has made it particularly urgent to improve the risk management capabilities of the pharmaceutical industry's supply chain in a transitional period. Indepth exploration of the key factors affecting the financial credit risk of pharmaceutical companies' supply chain, and the construction of a high-accuracy forecast model is of great significance to the stability of the macroeconomy. Combining the characteristics of the pharmaceutical manufacturing industry, this paper builds a financial credit risk assessment system for the pharmaceutical supply chain. On the basis of Factor Analysis and Random Forest variable screening, the AdaBoost algorithm is used to build the prediction model. By comparing basic machine learning models such as SVM model, decision tree, logistic regression, Bayesian classifier, BP neural network, and integrated learning models such as Random Forest, Bagging meta-estimator, GBM, and XGBoost, the study found that the AdaBoost model has higher accuracy. And through the data forecast in 2020, the superiority and effectiveness of the model for credit risk assessment in the pharmaceutical industry are further verified. According to the prediction results, this paper finds that the epidemic has no obvious negative impact on pharmaceutical manufacturing enterprises and proposes suggestions from the perspectives of the government and enterprises for reference.

Object detection is one of the most basic and important tasks in the field of computer vision, and it is the foundation of high-level vision tasks such as behaviour recognition and human-computer interaction. With the development of deep learning technology, the accuracy and efficiency of target detection models have been greatly improved. Compared with traditional target detection algorithms, deep learning uses powerful hierarchical feature extraction and learning capabilities to make a breakthrough in the performance of target detection algorithms. At the same time, the emergence of large-scale data sets and the greatly improved computing power of graphics cards have also contributed to the vigorous development of this field. This article reviews the existing research results of target detection based on deep learning. This article first reviews the traditional target detection algorithm and its existing problems, then introduces the two-stage detection model and single-stage detection model under deep learning, and finally briefly summarizes the application scenarios of the target detection algorithm and summarizes the full text.

For the prediction of the U value of the heat transfer coefficient of vacuum glass, a new method is proposed in this paper. By constructing a prediction model of vacuum glass heat transfer coefficient based on extreme random forest and random forest algorithm, the prediction of U value of heat transfer coefficient is realized. This paper measures the excellence of the prediction model by using the MAE, MSE and 𝑅ଶ squared value parameters and plotting the observed curve between the predicted value and the actual value. Finally, the evaluation values of extreme forest are 0.0458, 0.0050, 0.9784, and the prediction curves are very consistent, which proves that the extreme forest prediction model has a good U value prediction ability. At the same time, by introducing the RMSE image curve, it is observed that compared with random forest, extreme forest has better generalization ability under smaller data. Aiming at the difficulty of collecting vacuum glass data sets, this paper introduces the feature importance analysis method, and the correlation between the temperature change rate and the heat transfer coefficient U is as high as 0.9882. It provides a new idea for further reducing the size of the dataset.

To improve the performance of strapdown inertial navigation system (SINS)/ Doppler velocity logs (DVL) integrated navigation system, a new Interacting Multiple Model Feedback Particle Filter (IMMFPF) approach is proposed under the condition of complex and changeable underwater environment. In this scheme, the Interacting Multiple Model uses to deal with the model switch, while FPF approach takes each particle as a controllable stochastic system. By finding the solution about the Euler-Lagrange boundary value, the optimal feedback control law of Kullback-Leibler distance is obtained, and then the state of the particle itself is corrected and updated. Both simulations and field test illustrate that proposed IMMFPF approach can achieve a more accurate solution and make a more correct and effective response for SINS/DVL integrated navigation system.

Nowadays, falls are a common problem for older adults, which leads to injuries and decreased quality of life. According to the Chinese Center for Disease Control and Prevention, falls have become the leading cause of injuries to people over 65. Therefore, how to detect falls quickly becomes very important. In this paper, we assess the performance of five different video classification algorithms, namely, Efficient Convolutional Network for Online Video Understanding (ECO), C3D network, Temporal Segment Network (TSN), NeXtVLAD, Convolutional two-stream Network. Finally, we find that ECO and TSN algorithms perform better, and their accuracy is around 95%. However, ECO has a shorter running time and a faster response, so the ECO algorithm performs best in detecting falls.