Abstract:With the development of knowledge graphs, utilizing given knowledge graph data to automatically obtain answers to human natural language questions has become popular in recent years. QA systems such as Siri and Xiao Ai have been widely used. Thanks to the introduction of deep learning, breakthroughs have been made in various sub-projects in this field, but there are still difficulties that need to be overcome, such as multi-hop reasoning and strategy combination. Therefore, starting from the mainstream construction method, this study summarizes the current research status and challenges in this field, which can not only help researchers to efficiently carry out research in this field but also help researchers in different industries to develop industry-related QA systems to improve productivity.

Abstract:For further research promotion on the trusted transmission of national network data, this paper sorts out the development trend of research on source address verification to analyze its research status, development trend, and research focuses in China. Taking the literature on source address verification from the CNKI database as the research data source, this study adopts the methods of bibliometrics and mapping knowledge domains and uses CiteSpace, a visual tool, to carry out information statistics, co-citation statistics, and clustering analysis on the research samples. Then, these data are used to draw the interannual variation map of the literature in this research field and the knowledge map of co-occurrence clustering and time-series distribution for further scientific analysis. The study shows that the research on source address verification in China tends to develop dynamically and the trend is stable and good. Core research strength is headed by Professor Wu Jianping with Bi Jun and Xu Ke as important research experts. Research institutions are headed by Tsinghua University along with PLA Information Engineering University, and Chinese Academy of Sciences as important participants. Next generation Internet and software-defined networks are important emerging research focuses, which reflect the future research direction and the development trend of source address verification.

Abstract:As the most active research branch of artificial intelligence, deep learning has achieved significant improvements in fields such as computer vision, natural language processing, and speech recognition, and its applications in healthcare have also become popular in recent years. Deep learning has made achievements in medical image and signal processing, computer-aided detection and diagnosis, clinical decision support, medical information mining and retrieval, etc., which shows great application prospects. While introducing the principles of deep learning and common deep neural networks, this study gives a systematic and comprehensive introduction to the typical applications and latest research progress of deep learning based on clinical practice and published papers. Moreover, the discussion of difficulties and challenges of deep learning in modern healthcare is included in this paper, and some solutions or ideas are also provided correspondingly.

Abstract:With the rapid development of e-commerce platforms and new digital media services, the scale of network data continues to grow and data types are diversified. The mining of valuable information from large-scale data has become a huge challenge for information technology. Recommendation systems can alleviate the “information overload” problem, explore the potential value of data, push personalized information to users in need, and improve information utilization. The combination of the representational capabilities of deep learning and recommendation systems helps to dig deeper into user needs and provide accurate personalized recommendation services. This study analyzes the advantages and disadvantages of traditional recommendation algorithms, summarizes the research progress of deep learning technology in recommendation systems, and probes into the future development directions of intelligent recommendation systems.

Abstract:Segmentation of brain tumors in magnetic resonance imaging (MRI) is a difficult job due to the delineation data shortage, class imbalance, and significant differences among private databases. To solve those problems, we propose an adaptive local-global information learning (ALGIL) segmentation algorithm. This method only requires a small amount of delineation data, which solves the problem of traditional supervised learning depending on the amount of delineation data. The spatial-domain and frequency-domain information of the image is fused to convert the image from the spatial domain to the frequency domain through wavelet transform, and statistical features and texture features are extracted from the low-frequency and high-frequency sub-bands respectively. The limitations of traditional single-domain feature extraction are thereby resolved. With the ALGIL segmentation algorithm, we construct a similarity matrix by weighting the image according to the feature weights obtained through the random forest algorithm. Then, an exponential decay function is employed to adaptively adjust the degree of influence of the labeled samples on the algorithm. In this way, we solve the problem of poor segmentation results caused by the delineation data shortage. The experimental results of the proposed method on the public data set Brats2018 show that compared with other advanced models, the method has greatly improved the efficiency of image segmentation by enhancing various evaluation indicators and reducing the need for delineation samples. It provides a new idea for the automatic and accurate segmentation of brain gliomas.

Abstract:A steel bar is an indispensable structural material in the infrastructure industry, and accurate counting of steel bars is an essential link in both the steel-bar production process and the construction site. There are some problems in steel-bar bundles, such as dense end faces, non-uniform diameter scale, end-face boundary adhesion, fusion of end face and background, and end-face occlusion. To solve the above problems, this study proposes an improved YOLOv5 model framework to reduce the missed detection rate and the false detection rate of dense small targets. Considering the scarcity of the steel-bar end face dataset, the absence of a large public dataset in this field, and the weak feature of the steel-bar end face, we built a steel-bar end face dataset with the semi-automatic labeling method for dataset labeling and the data enhancement algorithm for dataset expansion. Moreover, the backbone network in YOLOv5 was modified, and the spatial pyramid pooling (SPP) and the small target detection layer were added to obtain larger feature maps. The feature pyramid network (FPN) and path aggregation network (PAN) were used to fuse multi-scale feature images to improve the accuracy of dense small target detection. Several groups of control tests were designed based on the Data Fountain steel-bar stocktaking competition dataset and the self-built steel bar dataset. The experimental results show that the improved algorithm YOLOv5-P2 model has the best performance on the steel-bar end face detection, and the mean average precision (mAP) of the steel-bar end face reaches 99.9%. Compared with the mainstream algorithms of YOLOv3, YOLOv4, ScaledYOLOv4, and YOLOv5, the proposed model has its mAP increased by 9.6%, 7.9%, 7.0%, and 1.1%, respectively. When tested in the real environment of factories, the model has stable performance, and its detection accuracy is improved by 2.1% compared with the original model on the test dataset. The position modification of the SPP module in the backbone network of YOLOv5 and the adding of detection layers can all significantly improve the detection accuracy of dense small targets with better edge feature extraction of the steel-bar end face and an mAP of 99.9%.

Abstract:With the popularity of cloud storage, increasingly more files are stored in cloud storage servers rather than in users’ computers, which makes users lose absolute control over the data and data security difficult to guarantee. To solve this problem, this study proposes a secure cloud storage system. It is implemented in the user model and can be directly set up on the computer file system, with low requirements for computer hardware and software. Functions such as end-to-end data encryption protection, integrity check, and access control are provided through the block encryption algorithm and the design of Merkle-B+ tree. The system is simple to use and completely transparent to users, with a reduced user threshold. The test results of this system show that when it is mounted on the network file system (NFS), its input/output (I/O) performance in a large file environment is reduced by about 5%, which indicates that the system has good performance in addition to ensured user data security and system ease of use.

Abstract:In recent years, the rapid development of heavy industry has exacerbated the deterioration of air quality, and environmental governance has become increasingly important. However, most of the existing air quality assessment systems at home and abroad have single forms, low accuracy and limited assessment ranges. In other words, they cannot accurately display the air quality situation in a diversified way. This study develops an air quality assessment system which integrates data collection, standardized processing and air quality assessment. The system has the following advantages. It obtains data from air quality monitoring websites with Web crawlers and enables data extraction, data cleaning, data unit conversion, pollutant classification and data processing with a variety of standardization methods to ensure the accuracy of air quality assessment. It has numerous assessment modes, involving the assessment by hour, day, month, monitoring station location and pollutant type, which addresses the single forms and inaccurate assessment of most air quality assessment systems. It can provide real-time air quality information for users and accurate data preprocessing results for researchers. With stable and reliable operation, friendly interface and rich functions, the system can meet the need for comprehensive air information management and evaluation.

Abstract:The C/S mode is an early architecture for developing Web servers, which is complex, costly, and lacking in generality. The B/S mode makes up for the disadvantages of complex use and high cost by concentrating system functions in the server, but it does not conduct in-depth testing and research on the environment and data transmission rate of the user end, which brings too many external visits to the Web page and puts pressure on the server. Given the limitations of the traditional C/S mode and B/S mode, this study introduces the idea of Docker container development and integrates it with Nginx and Flask methods respectively to construct the DoNginx mode and the DoFlask mode, which develop the Web server by modifying and establishing mirror images. The two modes combine the advantages of Docker’s lightweight, Nginx’s low consumption, and Flask’s stability to realize mode optimization. The CPU, integrity, and throughput performance tests are designed to make a comprehensive comparison with traditional B/S mode. Experiments show that the DoNginx mode has a high resource utilization rate, and the DoFlask mode has stronger environmental compatibility and reliability. Both modes are superior to the traditional B/S mode in the above aspects, boasting great contributions to the framework design and good experimental performance.

Abstract:The quality inspection after assembly of automotive interior parts is an important stage of assembly and an important guarantee for ensuring a high pass rate of interior parts assembly. The target detection hardware platform is built with low-power and high-performance NVIDIA development boards, and the Faster RCNN and YOLOv5 models are compared, and the YOLOv5 model, which has a better detection effect on small targets, is used to train the data collected by industrial cameras. The test results show that the accuracy of detecting 13 features of automobile interior fittings is as high as 95%, which realizes the efficient and accurate discrimination of automobile interior fittings and provides reliable auxiliary means for the assembly work of automobile interior fittings.

Abstract:For pulmonary nodules detected in computed tomography (CT) images, it is necessary to automatically determine whether they are at the risk of canceration. This study proposes a multitask learning model based on the attention mechanism. Different from most existing research methods which only distinguish between the benignity and malignancy of nodules, the proposed model also assesses and outputs the semantic features related to the benignity and malignancy of nodules. The assessment of nine nodule features (subtlety, lobulation, spiculation, sphericity, margin, texture, calcification, diameter, and malignancy) and the sharing of internal characteristics are conducted at the same time to improve the performance of each subtask. The vision transformer (ViT) model is selected as the multitask shared feature extraction layer, and the whole model uses the dynamic weighted average method to optimize the Loss function of each subtask. Experiments on the LUNA16 dataset show that the proposed learning framework can improve the risk assessment of pulmonary nodule canceration and that the assessment of other semantic features can also enhance the interpretability of the results.

Abstract:In view of the current waste of agricultural irrigation water, an intelligent irrigation system with a Raspberry Pi as the main control device and a cloud platform for monitoring is proposed. This study, with the crop as the experimental object, analyzes the characteristics of the fuzzy control algorithm, Raspberry Pi 4B, sensor technology, HTTP protocol, MySQL database, and OneNET cloud platform and designs an intelligent irrigation system that can automatically irrigate according to environmental conditions. The hardware part of the system can complete the collection of air temperature and humidity and soil humidity. In terms of the software part, considering the low development cost of cloud platform technology that can completely replace the traditional Web in function, applications are developed on the cloud platform, and they can control irrigation and upload the data to the OneNET cloud platform. Meanwhile, they can also import the environmental data into the fuzzy controller to make the irrigation strategy more scientific. In irrigation experimental tests, the maximum irrigation error of the system does not exceed 2%, which means the system can solve the problem of irrigation water waste and therefore has broad application prospects.

Abstract:Given that the traditional data process system cannot display the data process dynamically and drag and edit the operators in the data process, this paper proposes a data process visualization scheme based on scalable vector graphics (SVG) and Vue and gives the implementation process. With the high scalability and independence from resolution of SVG and the two-way data binding of Vue, the scheme can drag and drop the data process to create or edit it and thereby bring a dynamic interaction experience to users. The scheme, with high scalability, can also improve development efficiency and deliver favorable performance.

Abstract:The new generation of smart grids has greatly improved the security and reliability of power grids, which relies on smart meters to send data every 15 minutes. However, this may expose the privacy of users and also requires a huge computation cost. As a result, data aggregation technology is introduced. Most of the existing aggregation schemes are time-consuming and the system cannot run normally when its meter fails. In response to the above problems, this paper proposes an efficient data aggregation scheme with fault tolerance in smart grids. Specifically, it uses the improved symmetric homomorphic encryption technology to be lightweight. It can resist collusion attacks while supporting fault tolerance. The security requirements analysis shows that the scheme is secure, and the performance evaluation reflects the efficiency of the scheme, which fits the smart meters with limited resources.

Abstract:Given the low detection rates of small targets, low detection accuracy in complex scenes, and delayed detection in fire detection, an improved You Look Only Once v3 (YOLOv3)-based fire detection algorithm is proposed. Firstly, an improved K-means clustering algorithm is used to retrieve anchors that are more in line with the sizes of the flames and smoke. Secondly, spatial pyramid pooling is added after the Darknet-53, which improves the network receptive field and enhances the detection ability of the network on small-scale targets. Thirdly, the loss function is improved through complete intersection over union (CIoU), and the convergence of the loss function is sped up by taking into consideration the correlations of the center with the width and height coordinates when calculating the coordinate error. Finally, mosaic data enhancement is employed to enrich the background of the object to be detected, and the improved algorithm is trained and tested on a self-made data set. The experimental results show that compared with the YOLOv3 algorithm, the improved algorithm improves the flame AP from 94% to 98%, increases the smoke AP from 82% to 94%, and promotes the average detection speed from 31 fps to 43 fps. Compared with the Faster R-CNN, SDD , and other algorithms, it also has a higher mAP and a faster detection speed. Therefore, the improved algorithm is more effective in fire warning.

Abstract:Rib fracture is a common disease in clinical medicine. However, the diagnosis of fracture with the manual method is heavy in workload and difficult. To help doctors to reduce workload and improve detection sensitivity , we present a rib fracture classification algorithm based on a rib fracture network (RF-Net). Firstly, generative adversarial networks are used to generate synthetic medical images to enlarge the data size. Then, these data are input into our RF-Net to yield classification results. The data augmentation method can ease the overfitting phenomenon and improve the model training. In RF-Net, we use RF-block to replace the ordinary depth wise separable convolution , which can extract multi-scale features to strengthen the feature extraction ability of the entire network. Furthermore, considering the high requirement for fast speed, we apply the compression strategy to optimize some high-dimension modules to decrease the computation cost. The comparison with the existing deep learning models demonstrates that our method achieves the best result in multiple indicators, including accuracy, the area under the curve (AUC), sensitivity, and specificity. Besides, ablation experiments are conducted to verify the robustness of the algorithm and the effectiveness of each module. Finally, the results show that our method can classify the disease more accurately and faster than existing state-of-the-art approaches, which can provide a reliable basis for diagnosis.

Abstract:Falls are the first cause of injury-related deaths in people over the age of 65. A gait feature extraction method based on Kinect 3D skeleton data is proposed. This method can assess and predict the falls risks of the elderly according to the personalized features of the individual information of the subjects. The falls risks are divided into two classes: high falls risks and low falls risks. Considering the cost of data collection, the novelty detection model is used to train and access the feature data on an unbalanced data set. The experimental results show that the accuracy of one-class support vector machine (OC-SVM) detection is 86.96% and the F1-score is 88.55%, which means the proposed method can effectively distinguish subjects with low falls risks from those with high falls risks. These results also demonstrate the potential of predicting the falls risks of the elderly with Kinect 3D skeleton data.

Abstract:To improve the pedestrian detection performance, this study proposes a pedestrian detection algorithm based on improved YOLOv4 by combining SqueezeNet, attention mechanism, dilated convolution and Inception structure. An attention module named D-CBAM is proposed which is combined with dilated convolution. It is introduced to the feature enhancement part to select useful information from the extracted features. The residual connection is also used in this part to enhance feature reusability. In addition, an Inception-fire module is proposed by combining the “squeeze-expand” structure of SqueezeNet and the multi-scale convolution kernel structure of Inception, which replaces the continuous convolution layer in the network. Increasing the width of the network improves the performance of the algorithm and reduces network parameters. According to the characteristics of pedestrian detection and focal loss, the loss function is improved. The detection ability is enhanced through the addition of weights to the positive and negative samples and the hard and easy samples respectively and the strengthening of the training on positive samples and hard samples. The detection accuracy of the improved YOLO algorithm on INRIA person data set can reach 94.95%, which is 4.25% higher than that of YOLOv4. The parameters of the model are reduced by 36.35%, and the detection speed is improved by 13.54%. In short, the improved algorithm shows better performance in pedestrian detection than YOLOv4.

Abstract:Sleep staging is the basis of sleep data analysis. Given the dependence on manual extraction, the inefficiency of manual classification, and the inaccuracy of automatic sleep staging of current sleep staging methods, this paper proposes a method that combines two deep-learning neural networks, namely the convolutional neural network (CNN) and the bidirectional long-short memory neural network (BiLSTM), and uses electroencephalogram (EEG) data to conduct automatic sleep staging. This algorithm can extractmelspectrograms toobtain the original EEG dataand uses CNN and BiLSTM to extractfeatures in the time domain and the frequency domain. CNN can extract the high-level features of sleep signals, and BiLSTM can improvethe accuracy of automatic sleep staging when combinedwith the correlation of sleep data of different stages. The experimental results show that the proposed methodachievesan average accuracy of 89.0% in the three-state sleep staging task on the Sleep-EDF dataset. Compared with the traditional staging model based on statistical rules, this model is simpler, more accurate, and more efficient and has better generalization performance. The proposed algorithm is suitable for nonlinear, unstable, and non-stationary EEG data and effectively improves the accuracy of the results of the automatic sleep staging model. It possesses practical value in modern sleep medicine, sleep disorders, and other research.

Abstract:The traditional long-term correlation filter uses a single feature and cannot capture the target again after tracking failure. Considering this, the paper proposes a multi-feature fusion long-term target tracking algorithm combined with deep learning. On the basis of the long-term correlation tracking (LTC) algorithm, the proposed algorithm uses multi-feature fusion to join together the local binary pattern feature, the improved directional gradient histogram feature, and the color feature to promote the robustness of the tracking algorithm. Since the LCT algorithm adopts a random fern classifier to recheck the target, which has a limited detection range and low rechecking accuracy, the deep learning-based twin network instance search (SINT) method is employed to recheck the global image. The experiment in this paper is carried out on the OTC100 dataset, and the results show that compared with the LCT algorithm, the proposed algorithm has improved the range accuracy and the success rate by 13% and 10.3% respectively.

Abstract:Image super-resolution reconstruction technology can improve image resolution, which plays an important role in medical, military and other fields. The traditional super-resolution generative adversarial network (SRGAN) algorithm for image super-resolution reconstruction has a slow training convergence speed, and excessive sharpening of high-frequency texture leads to distortion of some details, which affects the quality of reconstructed images. To address these problems, the generator network and loss function of the traditional SRGAN model are improved for image super-resolution reconstruction. The sparse residual dense network (SRDN) is used instead of the traditional SRResNet as the generator network to fully utilize low-resolution image features. Meanwhile, the sparse connection method of SRDN and the depthwise separable convolution are used to reduce the number of model parameters. In addition, a joint perceptual loss of fused the low-frequency features and high-frequency features of VGG is proposed to improve the network’s perceptual loss function by combining with the mean square error loss. Tested on the Set5, Set14, and BSD100 data sets, the results show that the improved SRGAN algorithm outperforms the traditional SRGAN algorithm in three evaluation indexes, namely, peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and mean option score (MOS), and the details of the reconstructed images are clearer. The improved SRGAN algorithm shows better overall robustness and comprehensive performance.

Abstract:This study proposes a wrapper method based on a quantum-inspired evolutionary algorithm for feature selection in supervised classification. Firstly, it analyzes the shortcoming of excessively preferring classification accuracy in existing subset evaluation methods and then puts forward two new subset evaluation methods respectively based on a fixed threshold and a statistical test. Second, some improvements are made to the evolutionary strategy of the quantum-inspired evolutionary algorithm. More specifically, its whole evolutionary process is divided into two phases, in which individual and global extrems are selected as the evolutionary target of population respectively. On this basis, a feature selection algorithm is designed in accordance with the general wrapper framework. Finally, 15 UCI datasets are used to validate the effectiveness of the subset evaluation methods and the evolutionary strategy, as well as the superiority of the proposed method over other 6 feature selection methods. The results show that the new wrapper method achieves similar or even better classification accuracy in more than 80% of the datasets and selects feature subset with less number of features in 86.67% of the datasets.

Abstract:According to the nonlinear characteristics of battle damage assessment (BDA), this study proposes an equipment BDA method based on the belief rule base (BRB) and evidential reasoning (ER) in view of the multi-source and uncertain battle damage data in a complicated battlefield. Firstly, through the analysis of factors affecting BDA, a new BRB-ER model integrating multiple characteristics is presented. Secondly, to solve the problem of inaccurate parameters in the initial BRB of the traditional expert system, we use the local particle swarm optimization algorithm to optimize the initial parameters of the model and thus improve the accuracy of BDA. Finally, a battle damage test is taken as an example to verify and compare methods for equipment BDA based on the reasoning of BRB. The results show that the proposed method can effectively evaluate the equipment BDA and provide assistant support for a commander to make battlefield maintenance decisions.

Abstract:Complete and high-precision temperature observation data are important input parameters for agro-meteorological disaster monitoring and ecosystem simulation. Due to the limitation of meteorological field observation conditions, missing meteorological observation data is common. In response, interpolation becomes a necessary processing step before meteorological data application. In this paper, we construct a new deep learning model for interpolation of missing temperature data, which is employed to interpolate the missing half-hour temperature observations with high accuracy together with the low-frequency manual temperature observations at the same location. The deep learning model has a sequence-to-sequence deep learning structure based on the coding-decoding structure. A bidirectional LSTM-I (BiLSTM-I) network is used for the coding layer of the model, and an LSTM decoding structure and a fully connected decoding structure are respectively adopted for the decoding layer. The experimental analysis results show that the designed BiLSTM-I deep learning method for temperature interpolation is better than other methods. It can meet the need forhigh-precision temperature data interpolation. Particularly, the BiLSTM-I model with the LSTM decoding structure has higher data interpolation precision. The generalization ability of the BiLSTM-I deep learning model is also explored, and the results show that the model is effective in data interpolation for different lengths of the temperature missing window.

Abstract:With the development of artificial intelligence technology, it has been widely used in life and gradually penetrated judicial proceedings. However, there is insufficient interpretability in practical applications and thus it cannot effectively assist trials. In light of the four-element theory used in criminal case trials according to the constitution of a crime, this paper addresses the above problem by designing an identification task of the four elements constituting a crime. Some constituent elements of crimes of theft are screened, and a data set of the constituent elements is constructed. Moreover, a constitutive elements identification model is developed on the basis of the pre-trained language model BERT and then tested on the data set constructed in this paper, with the identification accuracy reaching 93.54%. Constructing an auxiliary sentencing algorithm based on the constituent elements can improve the interpretability of the existing algorithm and more effectively assist judges in hearing cases.

Abstract:The algorithm of the tread winding system configured in the open-loop control system for tread winding production features poor performance, and thus this study proposes a tread winding simulation algorithm based on the NURBS curve and verifies it with an example. At present, the tread winding simulation software of the open-loop control system for tire tread winding mostly uses straight lines and arcs for curve profiling in the winding process. The software’s built-in algorithm leads to a large deviation between the wound tread and the ideal tread and has a low winding efficiency. Therefore, this study uses the NURBS curve for curve profiling in the winding process and proposes a tire tread winding simulation algorithm based on the NURBS curve. The verification results show that the tread shape wound by the designed tread winding simulation algorithm meets the national standards and enterprise requirements and boasts good consistency and high production efficiency.

Abstract:The existing parking lot classification methods are exposed to problems of low-level automation and high equipment and deployment costs, and the existing detection algorithms have low recall rates and poor detection accuracy. To solve these problems, this study proposes a vision-based parking space detection and classification algorithm to improve the utilization efficiency of parking lots. First, parking spaces are detected to help build a parking space table andincrementally expand the parking space classification model dataset. Then, the test dataset is used to train the support vector machine (SVM) model for parking space classification. Finally, real-time judgment of the parking space conditions is made on every parking space based on the surveillance video data. The experimental results show that under different lighting conditions, the recall rate of the line detection of parking spaces is above 94%, and the accuracy of the parking space classification model is above 95%. The algorithm boasts a high degree of automation, good accuracy, simple deployment, and high application value.

Abstract:In view of the fact that the cache data cleaning of spark checkpoint needs to be cleaned by the programmer after the job is completed, which may lead to memory accumulation of the failure data. This study analyzes the execution mechanism of checkpoint, deduces that the checkpoint cache cleaning method is not extensible with the increase of the number of check points. The matching degree between checkpoint cache and memory slot is measured by using the utility entropy model of checkpoint cache. The optimal checkpoint cache cleaning time is derived by using the principle of best utility matching. The PCC strategy based on utility entropy optimizes memory resources by making the checkpoint cache clean-up time approximately equal to the time when the checkpoint is written to HDFS. The experimental results show that in the multi-job execution environment based on fair scheduling, with the increase of the number of checkpoints, the execution efficiency of the unoptimized program becomes worse. After using PCC strategy, the program execution time, power consumption and GC time can be reduced by 10.1%, 9.5% and 19.5%, respectively. Effectively improve the efficiency of multi-checkpoint program execution.

Abstract:To solve the problem of target loss caused by the large moving distance between two video frames during maneuvering target tracking in a video, this study proposes a kernel correlation target tracking algorithm based on one-step position prediction. Firstly, the gray scale and color feature of the target region in the current frame are extracted, and the gradient and color histogram operations of the gray scale feature and color feature are carried out respectively to obtain the FHOG feature vector and color histogram. Then, according to the color histogram, particle filter is introduced to predict the target position in the next frame. Further, the search area is determined with the predicted target position at the core. Finally, the estimated value of the current target position is corrected depending on the FHOG feature and the kernel correlation filtering in the search area. On this basis, the deformation and ambiguity generated during the movement of the maneuvering target are dealt with by the combination of the zero intercept and the average peak-to-correlation energy. Experimental results show that the proposed method can effectively improve the recognition rate in the UA-DETRAC dataset, and the success rate and accuracy of the proposed method are 11.96% and 9.6% higher than those of the standard kernel correlation filter algorithm.

Abstract:To solve the problem of the belated emergency rescue against drilling accidents caused by the long journey time of rescue vehicles, this paper proposes an improved ant colony algorithm for the path planning of drilling rescue vehicles. Firstly, in view of the deficiencies that the basic ant colony algorithm tends to fall into the local optimum, and the transition probability is calculated only depending on pheromone content and path length without the consideration of external factors affecting road traffic in the actual road network, the paper improves the basic ant colony algorithm by introducing path weight factors and optimizing path selection strategies. Then, the improved ant colony algorithm is employed to establish a path planning model for rescue vehicles with the least time as the goal. Finally, simulation experiments and practical application tests are carried out on the path planning of rescue vehicles. The results of experiments and tests show that the proposed method can reasonably plan a global optimal rescue path, which thus effectively solves the path planning problem of drilling rescue vehicles.

Abstract:Given that the traditional model-based collaborative filtering recommendation algorithm fails to effectively utilize the attributes of users and items and the relationship structures among users and items, this study proposes a collaborative filtering recommendation algorithm based onrepresentation learning with graph attention networks. The algorithm uses the knowledge graph to represent the attribute features of the nodes and the relationship structures among the nodes. Then, representation learning of nodes with graph attention networks is performed on the homogeneous networks of users and items to obtain their network embedding feature representations. Finally, a neural matrix factorization model integrating network embedding is constructed to obtain the recommendation results.This paper conducts comparative experiments with related algorithms on the Movielens dataset. Experiments show that the proposed algorithm can optimize the recommendation performance of the model and improve the recommended recall rate HR@K and the normalized discounted cumulative gain NDCG@K.

Abstract:In the research of passive network device identification based on network traffic analysis, much high-dimensional data often appears in the network traffic data, and some of these features do not contribute much to device identification and even can seriously affect the classification results and performance. Therefore, this study proposes a network traffic feature selection algorithm FSSA that combines Filter and Wrapper approaches based on symmetric uncertainty (SU) and approximate Markov blanket (AMB). Specifically, the proposed method in this study first uses the SU algorithm to select the features with classification contributions for each category and remove irrelevant feature attributes. Then, the AMB algorithm is adopted to delete redundant features in the subset of candidate features. Finally, the Wrapper approach based on the C4.5 classification algorithm is employed to determine the final feature preference. The experimental results show that the accuracy of the features selected under this method for type identification of the network device operating system has been improved compared with classical feature selection methods, and the recall rate on small class data has also been raised.

Abstract:The differential evolution algorithm is limited in the optimization of the differential strategy, the DE/best/1 strategy has a poor global detection ability, and the weak local search ability of the DE/rand/1 strategy leads to the reduction in robustness and local optimal problems. In this study, the differential strategy is improved and the idea of neighborhood divide and conquer is added to improve the evolutionary efficiency. A differential evolution algorithm (TPSDE) based on two-stage mutation strategy with two populations is proposed. In the first stage, the advantages of the DE/best/1 strategy are employed to locally optimize the subpopulation area with completed neighborhood vector partition.In the second stage,the idea of the DE/rand/1 strategy is borrowed to achieve global optimization. Finally, the final variant individuals are obtained by weighting the vectors of the two stages, which avoids problems such as premature convergence and search stagnation. The simulation results of six test functions show that the TPSDE has significantly improved the convergence speed, optimization accuracy, and robustness.

Abstract:The shallow models and linear functions are usually utilized for data embedding in data representation learning aimed at multi-view clustering. This strategy, however, cannot effectively mine the rich data relationships among the multiple views. For better representation of the consistency and complementarity information among different views, a tensor graph convolution network for multi-view clustering (TGCNMC) is proposed in this study. This method splices the traditional plane graphs into tensor graphs and uses tensor graph convolution to learn the neighbor relationships of the data in each view. Then, inter-graph convolution is adopted to transfer information among multiple views and thereby to capture the synergistic effect among the data of multiple views and reveal the consistency and complementarity information in those data. Finally, the self-monitoring method is employed for data clustering. Extensive experiments are carried out on standard data sets and the corresponding clustering results are better than those of the existing methods, which indicates that this method can represent multi-view data comprehensively, mine the relationships among views effectively, and deal with downstream clustering tasks beneficially.

Abstract:The current sentence similarity calculation method does not consider the multi-attributes of the keywords in the sentence and cannot better measure the sentence similarity. Therefore, this study proposes a sentence similarity calculation method based on multi-attribute fusion, considering the sentence structure and the attributes contained. First, this method extracts the attributes of the sentence including the word frequency, word order, part of speech, and sentence length. Next, the analytic hierarchy process (AHP) is used to calculate the weight of each attribute and verify the rationality of the weight, and then the weighted fusion of the similarity of the four attributes is conducted. This proposed calculation method for multi-attribute sentence similarity is tested on the constructed dataset to verify its reliability and feasibility, and it is compared with other traditional methods in recall rates, accuracy rates, and normalized F-metric values.The results show that this method has balanced recall and accuracy rates and a high F-measure value of 83.57%.

Abstract:Software-defined networking (SDN) has been gradually applied to tactical mobile ad-hoc network (MANET) routing research recently to provide better quality of service (QoS) than that of traditional MANET routing protocols for the increasingly abundant tactical maneuvering tasks. However, due to the strong variability and distributed structure of MANET, there are many problems in the application of SDN technology to MANET routing that need to be solved urgently. Starting from the status quo of tactical MANET routing, this paper explains the traditional MANET routing protocols and SDN technology and summarizes the existing MANET routing methods combined with SDN. Then, problems that need to be solved in the application of SDN to tactical MANET routing are discussed around the routing process, including the control layer reliability, flow table issuing time, control overhead, mixed network elements, and scalability. Corresponding methods are expounded from the perspective of the SDN structure. Finally, the issue of future research directions is discussed.

Abstract:With the development of technology, machine carving is increasingly used in the handicraft carving industry including olive-nut carving. We propose a method of model editing and tool path planning with automatic generation of 3D models suitable for shape-following olive-nut carving and rapid production of tool paths for olive-nut machine carving. We edit the existing 3D model and acquire a new 3D model suitable for shape-following olive-nut carving through rigid alignment based on the oriented bounding box, mesh deformation based on an improved radial basis function, and as-rigid-as-possible mesh deformation. Then, we slice the new mesh model and obtain the four-axis three-linkage tool path quickly generated according to the preset tool information. The experiments prove that the mesh models and tool paths suitable for shape-following olive-nut carving can be quickly obtained and favorable carving effects can be achieved by the proposed method.

Abstract:The prediction of fishing conditions is to predict the locations of fish schools and the abundance of fish in those areas. With knowledge of future fishing conditions, managers can formulate effective strategies and fishermen can cut down their resource consumption in the fishing process. This study starts with the remote sensing data of the marine environment and automatic identification system (AIS) fishing vessel trajectory data, analyzes the distribution of fish schools, and predicts future fishing conditions. According to different operation methods, fishing vessels can be divided into many types, such as purse seine, gillnet, trawl, and stow net types. It is of great significance to predict the future operation areas of fishing vessels equipped with different fishing gears and carry out fine management. The traditional single-task learning can achieve individual predictions for each fishing gear, but it cannot capture the interaction of various fishing gears. Therefore, this study proposes a multi-task prediction method based on a spatiotemporal neural network of ocean remote sensing data and AIS fishing vessel trajectory data. This method can capture the interaction of the fishing gears in addition to conducting separate predictions for each fishing gear. The prediction accuracy is further improved by embedding environmental remote sensing data such as ocean temperature and salinity into the model. Experiments are conducted on the data set of AIS fishing vessel trajectories in Zhejiang sea area, China, and the results prove the superiority of this method to the classical method and the latest onebased on ocean remote sensing and AIS trajectory inpredicting the distribution of fish schools.

Abstract:This work studies the driver response to variable message signs (VMS) based on the structural equation model. First, this study analyzes drivers’ travel habits and the reliability of VMSs. Then, the technology acceptance model is introduced to test the measurement model with two evaluation indexes of reliability and validity. 347 questionnaires are collected to form the dataset for the hypothetical testing of the proposed model. The results show that information reliability has a significantly positive impact on perceived usefulness and perceived ease of use, and the impact on the latter is greater than that on the former, while travel habits have no significant impact on perceived usefulness. Perceived usefulness and perceived ease of use have a significant positive impact on drivers’ behavior intention. This study can provide a theoretical basis and technical support for the traffic management department to scientifically release VMS information.

Abstract:Safety fences play an important role in power construction sites. However, violations of crossing the fence are widespread, causing great safety hazards to the construction sites. To intelligently supervise, this study proposes a Faster RCNN-based detection method for crossing fence that combines the object detection and the ideas of frame difference method. The proposed method first obtains the information of the fence location and human keypoints by object detection from the captured frames in the video and then recognizes the violations at the construction site with the frame difference method. The experiment results show that the method can effectively detect violations of crossing fences at construction sites and meet the real-time requirements.

Abstract:Exposed distress is one of the common diseases of cement pavements, which seriously affects pavement service life and driving safety. Therefore, it is very important to detect and repair exposed distress in time. Traditional manual detection methods are low in both detection accuracy and detection efficiency. In view of this, our study proposes a method for detecting exposed defects of cement pavements based on an improved RetinaNet model. Firstly, preprocessing operations such as filtering and denoising are carried out on the exposed distress images collected by manual and inspection vehicles, and the model training data set is constructed. Then the SE Net structure is embedded into the feature extraction network, and the feature pyramid network is improved. Finally, the detection of exposed distress of cement pavements is realized by the improved RetinaNet. The results show that the improved RetinaNet model enhances the detection accuracy of exposed distress by 4.9% compared with the original model, which reaches 98.9%. In comparison with Faster R-CNN, SSD and YOLOv3 methods, the model significantly improved the detection effect for the same test data.

Abstract:As logistics centers are an important hub of express transportation, their sorting efficiency has an impact on the delivery time of express packages. The coordinated sorting operations of multiple automatic guided vehicles (AGVs) can significantly improve the handling efficiency of logistics centers. This paper studies the non-conflict path planning in multi-AGV coordinated operations. The grid map is adopted to model the working environment, and a two-level path planning framework based on conflict search is proposed. In this framework, both the conflict search and constraint appending are achieved with the binary tree. When the upper level of this framework detects a conflict and adds the corresponding constraints, the lower level only needs to replan the paths of the AGVs related to the newly added constraints. The space-time A* algorithm is used to handle the path planning of a single AGV at the lower level. Furthermore, a conflict avoidance table is also introduced for avoiding the possible conflicts with existing paths of other AGVs. The simulation results demonstrate that the proposed multi-AGV path planning algorithm based on conflict search can solve various path conflicts.

Abstract:An improved KTBoost prediction model is proposed to address the low accuracy and poor fitting performance of the current KTBoost prediction model. First, the OGWO algorithm is put forward to solve the invalid iteration of the traditional gray wolf optimization (GWO) algorithm by using the arctangent function to optimize its convergence factor. Then, the OGWO algorithm is employed to optimize the hyperparameters in the KTBoost model, thereby improving the prediction accuracy of the model. Finally, the improved model and other prediction models are applied to traffic flow prediction scenarios for comparison to verify the feasibility of the model. The experimental results show that compared with the RBF model, random forest regression (RFR) model, KTBoost model, OGWO-RBF model, and OGWO-RFR model, the OGWO-KTBoost prediction model has better fitting performance and a better forecasting effect in practical application with its coefficient value of determination being 0.8265.

Abstract:In the traditional first-order hidden Markov model (HMM1), each state in the state sequence is assumed to be only related to the previous state. In this way, although the model learning and recognition algorithm can be simply and effectively deduced, a lot of information passed down from the above is lost. Therefore, in view of the traditional HMM1, a continuous sign language recognition method based on the second-order hidden Markov model (HMM2) is proposed to solve the problems of the difficulty and low accuracy of sign language recognition. In this method, a video of sign language is divided into several short videos by the sliding window algorithm, and the feature vectors of the short videos and word videos of sign language are obtained through the 3D convolution model. The relevant parameters of the HMM2 are thereby calculated, and continuous sign language recognition is achieved via the Viterbi algorithm. Experimental results show that the accuracy of sign language recognition based on the HMM2 is 88.6%, which is higher than that of the traditional HMM1.

Abstract:Given the unbalanced pedestrian attribute data, insufficient expression ability of pedestrian features, and poor robustness of current pedestrian attribute recognition methods, this study proposes a method based on local feature overlap and pedestrian attribute recognition. The network uses global and local branches to improve the overall feature expression ability of the network. In the local branch, the feature graph obtained is divided into several parts with the same size, and the loss of each attribute is calculated with the Focal loss function to solve the problem of pedestrian attribute imbalance. Finally, the optimal loss of each attribute selected by voting and the ID loss calculated through global features are taken as the total loss of the model. The proposed method is tested on Market-1501_attribute and DukeMTMC-attribute pedestrian attribute datasets, and the experimental results show that this method can effectively improve the accuracy of pedestrian re-recognition.

Abstract:Multi-candidate electronic voting schemes play an important role in elections. Full privacy implies privacy protection for both voters and candidates, which is an important property of secure electronic voting schemes. A “k-out-of-m” electronic voting scheme based on secure multi-party computation is proposed in the study. In the scheme, voters’ willingness is mapped into the form of an array. Combined with the ElGamal homomorphic encryption system, the scheme outputs election results through the interactive computing of voters and candidates in a semi-honest model, which achieves full privacy and does not require the participation of a third-party vote-counting agency. Furthermore, the number of dissenting votes is taken into account for the first time in order to avoid controversial election results.