Abstract:Mutation Based Fault Localization (MBFL) is a recently proposed approach with the advantage of high fault localization accuracy but limited to actual use in industry area since its huge mutation execution cost. Researchers improve MBFL’s execution efficiency from three aspects: reducing the number of mutants, reducing the number of test cases, and optimizing the mutant’s execution process. The former two methods were studied a lot and showed promising results, but in terms of reducing the number of test cases during MBFL, the related studies were limited for losing the precision of fault localization. In this study, we propose an Information Entropy based Test Cases Reduction strategy (IETCR) for MBFL. IETCR first calculates the information entropy of test cases, then sorts them according to the information entropy. Finally, it selects a small number of valuable test cases to execute mutants. Empirical studies are conducted on 100 fault versions of 6 programs from the SIR repository. The results show that IETCR can reduce 56.3%–88.6% mutation execution cost while keeping almost the same fault localization accuracy with the original MBFL.

Abstract:Reinforcement learning (RL) is a research hotpot in the machine learning area, which is considering a process of agent-environment interaction, sequential decision making, and total reward maximization. Reinforcement learning is worthy of in-depth research and a wide range of applications in the real world, and represents a vital step toward the Artificial General Intelligence (AGI). In this survey, we review the research progress and development in the algorithms and applications for reinforcement learning. We start with a brief review of the principle of reinforcement learning, including Markov decision process, value function, and exploration v.s. exploitation. Next, we discuss the traditional RL algorithms, including value-based algorithms, policy-based algorithms, and Actor-Critic algorithms, and further discuss the frontiers of RL algorithms, including multi-agent reinforcement learning and meta reinforcement learning. Then, we sketch some successful RL applications in the fields of games, robotics, urban traffic, and business. Finally, we summarize briefly and prospect the development trends of reinforcement learning.

Abstract:The timely feedback and repair of bug is the basic guarantee for the long-term and healthy development of open source software. Facing a large number of open source software bug reports submitted every day, including many factors, many bug reports affect the effective judgment and repair process of defects due to missing or inaccurate descriptions. For the complex bug report information that needs to be filled in, the report submitter is also impossible to determine which attributes are critical and which need to be highlighted, which results in poor quality of submitted bug report. Based on the analysis of 28 features of 5 dimensions described in bug report in the comprehensive literature, the importance ranking and comparison of bug report features are carried out from two perspectives of inter dimension and multi feature within dimension. The results show that the F1 value and accuracy index of two dimensions, text feature and bug reporter experience feature, are the highest, and the features within each dimension also show different importance, therefore, it can guide the bug reporter to standardize the submission of bug report, or guide the repairer to determine and repair defects.

Abstract:The three-dimensional tree model is widely used in the fields of virtual geographic environment, three-dimensional city scenes, etc. However, due to the rich geometric details of trees, it is still a huge challenge to effectively render trees in large-scale forest scenes. In this study, a real-time simplification method for tree foliage based on Viewpoint Mutual Information (VMI) is proposed. In the preprocess, the whole tree is divided into parent-child branch and leaf nodes according to its own topological relationship. Then, the average importance of each leaf under multiple viewpoints is calculated based on the VMI value and the leaves are sorted based on it. The leaves of less importance are pruned first during the real-time simplification process. In the real-time simplification process, we propose a view-dependent simplification method which greatly reduces the number of primitives needed to be rendered. In order to improve the rendering performance in forest scenes, a variety of rendering optimization measures are taken to avoid unnecessary Level Of Detail (LOD) transition.

Abstract:In the scenario of Cellular-Vehicle to Everything (C-V2X) based on the evolution of cellular communication, a base station as a Multi-access Edge Computing (MEC) edge cache node can improve the efficiency of user data acquisition, but its cache capacity is limited. Therefore, accurate cache request content prediction in C-V2X has become an important issue that needs to be addressed. This article starts with the time-varying characteristics of file requests and uses the Simulation of Urban Mobility (SUMO) to analyze the traffic flow for actual urban scenario modeling. Secondly, collecting the traffic data of the actual website time-sharing classification, and pre-processing according to the traffic flow rules of each road section, and then the user request model is constructed and the law of the base station receiving data is revealed. The Long Short-Term Memory (LSTM) deep learning model trains and predicts the file requests that each base station will receive. The simulation results show that the root mean square error of the vanillaLSTM model in the entertainment data set is about 1.3 when the data set received by the base station is predicted under the file request formed by NetEase news popularity distribution and request interval distribution.

Abstract:Water is an indispensable source of human being and other living species, thus it has significant value of social economy and ecosystem to establish water quality prediction model. This study developed a W-LSTM time series model to predict water quality based on wavelet decomposition and LSTM. Daubechies5 (db5) wavelet was used to decompose water quality data series into high frequency and low frequency signals, and these signals were used as the inputs of LSTM model to train the model to predict water quality data. Four water quality indices (pH, DO, CODMn, and NH₃N) collected from the Wangjiaba River basin in Funan, Anhui Province, China were used to train, validate, and test the model. The training and prediction results of the model were compared with these results of the traditional LSTM neural network model. The results show that the proposed model is superior to the traditional LSTM model in a variety of evaluation indicators. It is proved that this method has higher prediction accuracy and generalization ability and it is a more effective modeling and prediction approach.

Abstract:Aiming at the problem of bridge disease detection, especially the crack detection with high degree of damage, combined with the previous bridge detection system, an improved bridge detection system was proposed in this study. The hardware of the improved system is the DJI M210-RTK Unmanned Aerial Vehicle (UAV), and the software consists of image data acquisition module, crack detection module, and a module of 3D model building. In this study, calculation function of crack length and width is added to the crack detection module, and the length of the crack is calculated by curve fitting after iteration, besides, skeleton method is used to calculate the width. In the experiment, by setting the flight path, scanning distance, shooting distance of the UAV and the sub-region number of the bridge pier to be tested in advance, 200 pictures of the bridge pier deck and the video data of the bridge were collected. By identifying the crack types of bridge deck, and calculating length and width of crack, it can make us having a more comprehensive understanding of crack information and the degree of damage, and manual measurement in later period can be reduced, besides, combined with Ubuntu 16.04 system, the 3D model can easily and intuitively display the general situation of the bridge with using Direct Sparse Odometry (DSO) to carry out bridge 3D modeling. The improved system is stable, the method saves time and effort, and has wide applicability, especially for the detection of some sea-crossing bridges and bridges with complex surrounding environments.

Abstract:In the field of artificial intelligence, problem understanding has always been one of the most important and difficult problems in the automatic solution of mathematical application problems. In this work, we systematically studied the problem of lacking commonsense knowledge in solving elementary mathematical problems, and constructed a commonsense knowledge base to assist computer in problem understanding. We collected the real and simulated questions of the classical probability for college entrance examination over the years as the research object. On the basis of in-depth analysis of topics, this study put forward the concept of commonsense knowledge definition, analyzed and classified its features, and stored commonsense knowledge with XML tags. On this basis, the commonsense knowledge base was applied in practice. The experimental results indicate that the commonsense knowledge base constructed in this study is very helpful for the automatic solution of classical probabilistic application problems.

Abstract:In view of the new scheduling problem under the condition of multi constraints in the new curriculum reform, this study proposes a new optimized genetic algorithm strategy, and constructs a set of trial run scheduling system in a middle school. The new system integrates the student selection module, student performance module, and student evaluation module. Compared with the traditional genetic algorithm, the conflict chromosome optimization strategy is proposed in this study for the first time; in the genetic algorithm, a new conflict chromosome operator is added, which improves the efficiency of course arrangement by 19.2%. Under the condition of adaptive mutation rate optimization, by adding the conflict chromosome, we can cut off the useless solution in the iterative process of the algorithm, which can not only ensure the search space of the solution, but also accelerate the convergence of the algorithm. In addition, the effect of students’ independent choice of subjects and division of classes on class arrangement is verified. The experiment shows that the efficiency of students’ division of classes according to the strategy of “combination of course selection” and the combination of class arrangement with teachers, classrooms, time, and other educational resources can be improved even more.

Abstract:Due to the high requirements for security and confidentiality of contract data and the potential security risks in the Internet, the current contract management system cannot fully utilize the advantages of the Internet to achieve online operations. Identity authentication and contract signing are technical difficulties in online operations. In order to realize the contract management system running on the public network server, and complete the company identity authentication and contract signing online, this study designs and implements a contract management system based on private chain technology. First, this study proposes the design concept of the company’s digital authentication center, which is used to manage the company’s digital information and participate in the company’s identity authentication; then it designs the implementation scheme of the contract management system and private chain system, and finally introduces the technical points in the implementation of the system. The analysis of experimental results shows that the management system can realize all the processes of contract management online, which improves the efficiency of contract management.

Abstract:Currently, most of vehicle structured information was obtained through multiple steps, which caused the problems such as fussy training, limited training data in each step, and the accumulation of error in processing. Therefore, multi-task learning was applied to union the structured information extraction in a single neural network, and shared feature extraction structure can help to reduce the error accumulation in various processing. A loss function of multi-task learning was put forward for end-to-end network training. For solving the training data limitation problem, a new dataset augmentation and combination approach was advanced. The experimental results on the dataset KITTI show that the mAP (mean of the Average Precision) of VSENet achieves 93.82%, and the processing speed can satisfy the real-time request. Compared with multi-step vehicle structure information extraction method, the proposed approach reduces 60% of average processing time and achieves similar or better performance. These results show that the proposed method has certain advancement and is effective.

Abstract:The Web visual development platform focuses on implementation of visualized Web applications. It not only integrates single sign-on, automatic graph generation, form designer, layout designer, device topology creation, and quickly develop functional components base on business, but also analyzes the business needs of power grid in various regions and further strengthens the functions of each component. The core of the platform realizes the functions of zero coding and visual development of Web application systems, and realizes the separation of business logic and code programming. At the same time, the platform combines flexible multi-component configurations with specific functional requirements for power grid scheduling in various regions. Through North Hebei and Fujian automation equipment development project practice, it effectively verifies that Web visual development platform can effectively improve the development efficiency of business personnel and respond to changes in demand in time, modular components can be flexible to customize the interface at the same time.

Abstract:The SF6 gas density meter is a pointer instrument arranged in the substation to measure the density state of the SF6 gas which is the arc extinguishing medium of circuit breaker. In order to solve the problem that the traditional pointer meter is difficult to apply in the substation automation management, this study proposes a remote meter reading system for substation SF6 gas density meter based on LoRa wireless communication technology and machine vision, and carries out the system’s overall design, hardware design, and software design. This design can achieve remote automatic management of the substation pointer SF6 gas density meter in a high efficiency, low power consumption, and low cost way. This system has clear functions, convenient layout, easy expansion, and low maintenance cost. It has a very important reference value for the application of substation automation management.

Abstract:At present, ZigBee, WiFi, and other communication technologies are widely used in the realization of communication architecture of smart agriculture system. Although these can basically achieve long-distance transmission, low-power consumption, and other requirements, there are still many shortcomings in such aspects as anti-interference and cost. To solve these problems, this study introduces a smart agriculture system based on LoRaWAN technology. The LoRa terminal node is designed by using STM32 MCU as the main controller and LoRa RF module as the data transceiver module, and the raspberry pi is used to build a LoRa gateway concentrator for data forwarding, we finally realize the monitoring of environmental of crop growth and the portable system management function by setting up the ChirpStack service and Flask Web application on the cloud server. The working principle and system design of the LoRaWAN application system are described in detail. The system has stable data transmission and strong anti-interference, also dramatically reduces development costs.

Abstract:Copy move forgery is a common way of digital image tampering, which has become an important research direction in the field of multimedia forensics in recent years. We propose a robust copy-move forgery detection algorithm. We construct wave function to extract keypoint evenly in the image, which can extract enough keypoint even in small or smooth areas. We introduce DBQ-LSH for feature matching, which greatly reduces time consumption. We propose a novel approach which use invariant moment LBP image to locate forged areas, even if the image is seriously attacked, the algorithm detection accuracy is still excellent. A large number of experimental results verify the reliability of proposed algorithm, in terms of validity and accuracy.

Abstract:In view of the discontinuity and uncertainty of moving objects in the non overlapping view of multiple cameras, a handover algorithm of moving pedestrian target based on deep learning is proposed. Firstly, a face feature extraction model is constructed based on deep convolution neural network, and the face feature extraction model is trained to obtain accurate face features. Then compare two common similarity measurement methods, choose a more suitable similarity measurement method to complete the optimal face matching process, and then improve the accuracy of face matching. Finally, the most matching face can be found by feature matching under different cameras to realize the handover of moving objects. Experiments show that the depth neural network can accurately extract the facial features of moving objects, achieve the accurate matching process of human faces, and effectively complete the task of moving object tracking under multi cameras.

Abstract:The clustering characteristics of high-dimensional data are usually difficult to observe directly. Constructing it into a complex network, the topological structure of the network nodes can reflect the relationship between samples. Community detection of nodes in the network can achieve more intuitive clustering of data. A low randomness label propagation clustering algorithm based on network community detection is proposed. First, the data set is constructed as a sparse fully connected network using the radius and nearest neighbor methods. Then, according to the similarity of the nodes, the node labels are preprocessed to make the similar nodes have the same labels. The influence value of the nodes is used to improve the label propagation process and reduce the randomness of label selection. Finally, based on the cohesion, the community is optimized and merged to improve the quality of the community. The experimental results on real data sets and artificial data sets show that the algorithm has better adaptability to all kinds of data.

Abstract:A shuffled frog leaping algorithm is proposed to overcome the defects of fall into local optimum easily and the lack of ability to solve discrete optimization problems when solving high-dimensional complex problems. In the proposed algorithm, using the perturbation coefficient to regulate the movement of individual frog distance, so as to better balance the global search and local development capabilities of the algorithm; using the spiral update position strategy to enable the algorithm to perform a more comprehensive and refined search near the optimal solution; using a random search strategy to improve the global search ability of the algorithm; using the 2-opt method to implement the global optimal solution mutation to increase the diversity of the population; the SFLA algorithm is discretized by the improved Sigmoid function. The optimization experiments are conducted on the 9 benchmark functions and oilfield measures planning. Simulation results show that the proposed DSFLA has a better search performance.

Abstract:Aiming at the image recognition of wear particles, such as sever sliding, fatigue spall, and laminar particles, image shape feature extraction method based on shape signature and bispectrum analysis was put forward. Firstly, according to four shape signature methods which are centroid distance function, cumulative angular function, farthest point distance, and triangle area representation, the two-dimensional wear particle images were converted to one-dimensional signal. Secondly, normalized bispectrum was got by carrying out bispectrum analysis on one-dimensional signal. At last, by calculating bispectral invariants according to bispectral integration and bispectral moment on normalized bispectral domain, 76-dimensional shape feature was got, which covered whole feature, angle change information, angular point information, and contour detail information of the shape. In order to evaluate the method, shape recognition ability experiment and anti-noise ability experiment were carried on MPEG-7 CE Shape-1 Part B dataset and Swedish leaf dataset. The experiment results demonstrates that the proposed method can enhance the recognition accuracy rate and anti-noise ability of bispectrum analysis.

Abstract:The gas load is affected by a variety of factors, which cause the trend of gas load changes to have greater complexity and more redundancy of eigenfactors. In order to solve this problem, the EEMD adaptive time-frequency localization analysis method is used to deal with the complexity of the gas load. The non-stationary gas load data is decomposed into stationary eigenmode components and residual terms. Mutual information analysis is added to the PCA to select the eigenvectors. The relationship between the features and the gas load value can be considered in the dimension reduction. The corresponding LSTM model outputs the final result. Using Shanghai gas data for verification, the experimental results prove that the method proposed in this study has a MAPE of 6.36%, which is lower than the error of other models.

Abstract:In order to save resources, reduce the electricity expense of the end user, reduce the cost of the power company, and satisfy the various cold and hot load demands of the users, a method of resource regulation and optimization under the platform area is proposed. By analyzing the typical scenario of layered distributed coordinated operation of resources under the platform area, a layered distributed coordinated optimization model was established with the dual objectives of the local agent layer and the power supply agent layer as the optimization objectives. Finally, the feasibility of the proposed method is verified by examples in different typical scenarios to provide technical support for the optimization of adjustable resources.

Abstract:Target tracking is a research hotspot in the field of machine vision. How to improve the tracking level in complex scenarios is a challenging problem. Previous studies have shown that how to use features effectively is the key to tracking. Therefore, a target tracking algorithm based on channel fusion features is proposed. Based on the multi-channel correlation filtering framework, the method introduces the feature channel weight, adjusts the weight according to the contribution of the channel to the response value, and then constructs the real-time feature combination. The algorithm can capture the state change of the target quickly and track the target effectively. In order to verify the effectiveness of the algorithm tracking, we test the algorithm performance on the open dataset OTB2015 and compare it with a variety of tracking algorithms. Experimental results show that the algorithm has sound tracking accuracy and success rate, and the overall performance is better than the compared algorithm.

Abstract:According to the current method of constructing Fractional Repetition Codes (FRC), it is found that most of them are distributed storage systems based on isomorphism, but the actual storage systems often need to satisfy the characteristics of heterogeneity. To this end, this study proposes two methods for constructing heterogeneous FRC. One is a heterogeneous FRC constructed based on matrix transformation. This method is used to construct an FRC with a repeatability of 2 and a heterogeneous storage capacity of nodes. Compared with the existing isomorphic FRC constructed by regular graphs, it has the advantages of lower computational complexity and more in line with the real storage system. In addition, this study also proposes a method of constructing FRC using adjustable rings, which FRC is constructed with a repeatability of 2 or 3, which can obtain the FRC of the node storage capacity is isomorphic, and can also get the heterogeneous FRC. Compared with the existing FRC, it is found that the FRC constructed in this study has heterogeneous characteristics in node storage capacity, good repair locality, and low computational complexity of the construction algorithm. It can select parameters in a wide range and the construction structure is simple and intuitive.

Abstract:To overcome the limitations of low efficiency of traditional face recognition, a novel method of face recognition based on Image Gradient Compensation pattern (IGC) is proposed. Firstly, gradient magnitude maps of a face image in four directions are calculated. Secondly, two gradient operators are produced by fusing the four gradients magnitude maps of a face image in multiple ways. Thirdly, the new gradient operators are used to compensate the original image and generate the IGC of the face image. Next, IGC feature maps are divided into several blocks, and the concatenated histogram calculated over all blocks is utilized as the feature descriptor of face recognition. Finally, Principal Component Analysis (PCA) is used to reduce the dimension of high-dimensional features. The recognition is performed by using the Support Vector Machine (SVM) classifier. Experimental results on YALE and CMU_PIE face databases validate that the algorithm in this study not only achieves high recognition rate, but also has excellent performance in computational efficiency.

Abstract:In mobile cognitive wireless sensor networks, the mobile characteristics of nodes will lead to continuous changes in network topology and uneven energy consumption of nodes. This study proposes an event-based clustering algorithm for mobile cognitive wireless sensor networks. The above problem is solved. The algorithm determines the qualified node and the standby node according to the pre-estimated dwell time in the communication area and adopts the direct clustering method to build the node by the moving direction, speed, and pre-estimated connection time of the node in the cluster. Clusters improve the stability of the cluster and ensure the minimum number of router hops. Compared with the simulation algorithms of mESAC, EACRP, and MNB, the algorithm has lower cluster energy consumption and better connectivity.

Abstract:In this study, a double fuzzy control algorithm for automatic light locking of intelligent curtain system is designed on the basis of traditional PID regulation, aiming at the requirements of anti-interference and robustness of intelligent curtain control system in practical use. Using the double fuzzy control to the input voltage and output speed of the DC motor, the system can automatically, steadily, and quickly control the curtain opening and closing according to the indoor light intensity, and then complete the accurate adjustment of indoor light intensity, and realize the light locking function of the intelligent curtain system. The mathematical model of the system was established by Simulink and fuzzy controller module in MATLAB to verify the feasibility of the double fuzzy control algorithm. The experimental results show that the algorithm can adjust the speed and steer the motor in real time according to the change of regional light conditions, and it has the characteristics of short adjustment time, high sensitivity, and high precision, and can achieve the optimal control of adjusting the indoor light intensity through curtains.

Abstract:Deep learning is widely used in different fields. However, a well-trained deep learning model may be easily disturbed and gives wrong results, which causes serious safety problems. In order to test the robustness of deep learning model, researchers attack the model by all kinds of adversarial examples. The generation method of black box adversarial examples with targets, which has sound practicability, becomes a hot issue. The difficulty of black box adversarial examples generation lies in how to improve the generation efficiency under the premise of the success rate of the attack. In order to solve this difficulty, this study proposes a new method of target adversarial sample generation based on fast boundary attack. This method includes two steps: sampling along the line and sampling on the sphere. The first step is completed by the one side half search to improve search efficiency. The second step is completed by random search with adaptive adjustment of search radius, which is used to improve the search scope. The feasibility of the algorithm is verified by experimental results of five groups of pictures.

Abstract:In order to solve the problem of low accuracy and speed of traditional SURF (Speeded-Up Robust Features) algorithm, an optimized image matching algorithm is proposed. A local two-dimensional entropy is introduced to characterize the uniqueness of feature points in the stage of feature point extracting. Some error points are eliminated by calculating the local two-dimensional entropy of the feature points and setting appropriate thresholds. The Euclidean distance is replaced by the Manhattan distance during the matching phase. The concept of the nearest neighbors and nearer neighbors is introduced. The first two points are extracted with the closest Manhattan points between the feature points in the template image and one in the image to be matched. If the ratio obtained by dividing the nearest distance by nearer distance is less than the threshold T, then this pair of matches is accepted to reduce mismatches. The experimental results show that the algorithm is superior to the traditional algorithm, and the accuracy is improved while the speed is also improved.

Abstract:The goal of multi-view clustering is to divide data exploiting the consistent and complementary information from various views. However, the ability to represent data varies from view to view, and some views may even contain a lot of redundant and noise information which not only cannot bring diverse information, but also affect the clustering performance. In this study, an adaptive weighted low-rank constrained multi-view subspace clustering algorithm is proposed, which construct the latent consensus low-rank matrix shared by each view and each view is given adaptively learned weights. An effective iterative optimization algorithm is proposed to optimize the model. Experimental results on five real data sets show the effectiveness of the proposed algorithm.

Abstract:Network anomaly detection technology has become the focus of research in the field of intrusion detection. However, because most of the current network anomaly detection remains at a single point of network anomaly detection, it cannot respond quickly and timely to joint anomaly attacks and malware that are constantly updated. In this study, a industrial control network anomaly detection algorithm based on graph neural network is proposed, which combines the network node’s own attributes and the information of neighbor nodes in the network topology to realize the network anomaly detection. First, each network node obtains a state vector that contains the feature information of the connected nodes and the interaction information between the nodes. Second, each node uses the fixed point theory to iteratively update the network. Thirdly, according to the node’s own information and neighbor node’s information, extract higher-level features through the neural network as the representation of the node. Finally, clustering is used to detect the abnormal behavior of industrial control network nodes. Experimental results show that the algorithm proposed in this study has high detection rate and high robustness.

Abstract:In the field of clinical medicine, image aided diagnosis has a high demand for the effect of medical view processing. To solve the problem of poor visual effect in medical image fusion, a medical image fusion algorithm based on sparse theory and fast finite shear transform is proposed, which improves the efficiency of medical image processing. Firstly, Fast Finite Shear wave Transform (FFST) is used to decompose the source image into high-frequency coefficients and low-frequency coefficients. Secondly, according to the different properties of high-frequency coefficients and low-frequency coefficients, different fusion strategies are provided, and the high-frequency coefficients are processed by the relative standard deviation comparison method. For the low-frequency coefficients with poor sparsity, K-SVD method is used to train, and a dictionary is obtained Finally, the high-frequency and low-frequency coefficients after fusion are fused into the medical image by FFST inverse transform. The experimental results show that the image fusion effect of the algorithm is good, especially in improving the image clarity and so on, which has sound practical value and application prospects.

Abstract:In order to improve the current collection ability of high-speed train and reduce the off-grid rate, this study designs the active controller of pantograph based on linear quadratic optimal control. Aiming at the problem of weight matrix Q and R in the linear quadratic optimal control, genetic algorithm is used to optimize. The objective function of the system is calculated by the dynamic performance index of the system and the optimal value of the weight matrix is obtained. And it solves the problem that the weight matrix is difficult to realize the global optimal in traditional linear quadratic optimal control by the experience design. Through the simulation and analysis of the change of catenary stiffness and the change of contact pressure parameters between pantograph and catenary at different speeds, the active controller designed in this study can reduce and control the fluctuation of contact pressure, and improve the dynamic performance index of pantograph and catenary system.

Abstract:Research on distributed storage and computing technology of spatial vector data is carried out. The method of quadtree grid coding to establish feature index is studied. HBase pre-partition optimization strategy is designed, a distributed storage model of spatial vector data is proposed. Based on MapReduce computing framework, the process of spatial data distributed computing and analysis is built. For the common application scenario of spatial overlay analysis and statistics, a large-scale data test is carried out. The results show that the scheme is effective.

Abstract:With the development of deep learning technology, the performance of natural scene text detection has been significantly improved. Nonetheless, two main challenges still exist: the first problem is the trade-off between speed and accuracy, and the second one is to model the arbitrary-shaped text instance. In this study, we propose a segmentation-based method to tackle arbitrary-shaped text detection efficiently and accurately. Specifically, we use a low computational-cost segmentation head and efficient post-processing. The segmentation head is made up of Feature Pyramid Enhancement Module (FPEM) and Feature Fusion Module (FFM). FPEM can introduce multi-level information to guide the better segmentation. FFM can integrate the features given by the FPEMs of different depths into a final feature for segmentation. We use a Differentiable Binarization (DB) module, which can perform the binarization process in a segmentation network. Optimized along with a DB module, a segmentation network can adaptively set the thresholds for binarization, which not only simplifies the post-processing but also enhances the performance of text detection. On the standard datasets ICDAR2015 and Total-Text, the method proposed in this study uses a lightweight backbone network such as ResNet18 to achieve comparable results in terms of speed and accuracy.

Abstract:In order to analyze the sentimental focus of the comment data from users of masks during the outbreak of virus, we extracted 143 330 comments about the purchase from Taobao users from March 1st to April 11th, 2020 by means of the Web Scraper of Google browser. To improve the accuracy of the sentimental estimation, each comment of the total 14 400 pieces was manually marked as positive or negative emotion on this data set. And then we used SnowNLP, the sentimental analysis model to train them. At last, the trained corpus was used for sentimental estimation. The overall sentiment of the comments was proved positive. On the basis of the daily emotional variation trend of users’ comments, the trend of local new cases (excluding overseas input) to some extent affects the overall change of their daily emotional trend. And the local fluctuation trend of domestic new cases (including overseas input) also affects that of the everyday emotional performance. After classifying the predicted comments, we found that users’ positive comments focused on the quality, packaging, price, and thickness of masks, while negative comments focused on the quality, packaging, smell, and whether the masks were for medical use.

Abstract:With the rapid development of the Internet and the increasing number of Internet businesses and users, more and more traditional monomer applications have chosen to split the business into multiple microservices for the convenience of expanding new business and increasing reusability, which is convenient for later management and expansion. However, it is very cumbersome to deploy multiple microservices on the cloud platform in traditional way, which consumes human and material resources. In order to realize agile development and rapid deployment, and reduce the time loss of team between development and operation and maintenance, this work studies the deployment of CI/CD pipeline service in the experimental environment of Kubernetes, a distributed container arrangement engine platform, so as to realize the automatic construction of code to service.

Abstract:With the wide application of message-oriented middleware in large-scale distributed systems, its security should be paid more attention. This study analyzes the security risks in the traditional shared response queue application of ActiveMQ, and it proposes a message-oriented middleware application mode based on random response queue. In this application mode, the response queue name is randomly generated by user client. Only the client and server know the random name. The random queue is very hidden, so it can ensure security application of the message queue. The basic framework, operation flow and safety analysis of the application mode are given. The operation performance is analyzed by theoretical calculation. The conclusion shows that the security application mode does not affect the operation performance of the system.