Abstract:With the large-scale application of deep learning in the field of object detection, the accuracy and speed of object detection technology have been rapidly improved, and it has been widely used in many fields, including pedestrian detection, face detection, text detection, traffic sign and signal light detection, and remote sensing image detection. This study reviews object detection technology based on the investigation of relevant domestic and foreign literature. First, the research status of object detection as well as the datasets and performance indicators for object detection algorithm tests are introduced. In this paper, two kinds of typical object detection algorithms with different architectures, namely two-stage object detection algorithms based on region proposals and one-stage object detection algorithms based on regression analysis, are described elaborately in their process architectures, performance effect, advantages, and disadvantages. In addition, some new object detection algorithms developed in recent years have been supplemented, and the experimental results and advantages and disadvantages of various algorithms on mainstream datasets are listed. Finally, some common application scenarios of object detection are specified, and future development trends are analyzed considering current research hotspots.

Abstract:Multi-robot coordination is one of the hot spots in future robot research, with dual-robot systems acting as a typical representative. According to the characteristics and common applications of dual-robot cooperative systems, this study introduces the research contents on these systems from three aspects, namely the establishment of dynamic models, trajectory planning and cooperative control. Moreover, it analyzes the technical vulnerability and difficulties in various aspects at present and points out the future development direction.

Abstract:Most of the current vehicle recognition methods rely on deep learning to directly input image data for training, thus obtaining a deep network. Due to the perspective distortion and scale change of an image, a large number of different types of data have to be used for training, without obtaining the vehicle-related physical information. To address the above problems, we propose a method of vehicle fine-grained recognition based on inverse projection space. First, the three-dimensional bounding boxes are constructed for vehicles under projection of a monocular camera by calibration information and geometric constraints. Second, the bounding boxes are unfolded to obtain normalized and standardized three-dimensional data in the inverse projection space. Finally, a deep convolutional network is introduced to obtain vehicle recognition results and its corresponding physical sizes of five common types of vehicles by training these standardized data. Experimental results show that, compared with traditional end-to-end vehicle recognition methods based on deep learning, the proposed method can effectively improve the accuracy of recognition while using less training data, and the three-dimensional physical sizes of vehicles can also be obtained simultaneously.

Abstract:With the increase in scale and complexity of software projects, a large number of bug reports are generated during the testing process, among which duplicate bug reports are widely present, reducing the efficiency of developers in fixing bugs. The prediction of duplicate bug report has become one of the popular research fields in recent years, and its efficiency and accuracy need to be improved. Therefore, this study puts forward a prediction method of duplicate bug reports based on semantic extension and continuous queries. Through the construction of a bug report index thesaurus based on the theme model, the semantic extension of query sequences is conducted. Then, the bug report retrieval algorithm based on the continuous query is adopted to narrow the index space and improve the prediction accuracy and efficiency. Experimental results show that compared with the traditional prediction method of duplicate bug reports, the proposed method reduces the index space of bug reports by more than 50%, improves the prediction effect by up to 33.6%, and shortens the retrieval time by 41%–73%.

Abstract:To achieve accurate prediction of settlement during tunnel construction, this study proposes a prediction method for settlement in construction tunnels through neural networks based on spatiotemporal feature region. This method effectively integrates multi-dimensional spatial characteristics and makes a reasonable prediction of the future evolution trend according to the current tunnel ground settlement. Taking the ground observation data at the Luanchuan end of Baijiazhuang Tunnel as an example, this study analyzes the prediction performance of the proposed method. The prediction results show that the sensing data of tunnel ground settlement is accurate and robust. The research can be applied to the monitoring and management process of actual tunnel construction.

Abstract:As scientific research is increasingly dependent on fast data transmission, the requirements for link resource planning and operation management of scientific research networks are more demanding. Considering the actual needs of scientific research networks, a good link traffic prediction model can help the network operators make fast decisions on link resource scheduling more effectively with the assistance of flexible network control technology such as SDN. The existing prediction model has ignored the current network traffic is more diversified and more complex in fine-grained features. This study proposes a new link traffic prediction model based on the improved LSTM model to meet the management needs of scientific research networks. Composed of AutoEncoder (AE), Bi-LSTM model, unidirectional LSTM model, and fully-connected layers, it can greatly improve the extraction ability of traffic features and better explore the dependent manners among data features at different time. The model is verified by using the associated node data of a link randomly selected from the real production environment of the national backbone network of Science and Technology Daily—CSTNet. The experimental results show that the prediction results of the model accord with the real change trend of traffic, and the residual between the predicted value and the observed value is small, which means the model can well fit the existing traffic of the scientific research network.

Abstract:The analysis and prediction modeling for the visual presentation of provincial big data concerning realty transaction and registration is of great significance for studying the layout trend of China’s urban-rural construction and regional economy. It shows the temporal and spatial evolution of urban construction and development indicators and supports scientific decision-making and macro-control. The prediction modeling of these economic activity data involves the understanding of the state evolution of things with complex factors and without clear mathematical expression. Thus, inspired by the successful applications of modern artificial intelligence-deep neural network technology in similar complex scenes, we intend to establish a macro visual prediction system of heat maps of provincial realty big data by related long short-term memory (LSTM) model and fully connected layer (FC) technology. The main system construction practice of this paper is that we utilize the big data from the legal business of realty which are accumulated in Guangdong Province (not including Dongsha Islands) over the years to implement the functions of modeling and predicting regional year-end geographic heat maps of two basic indicators, i.e., the number and the total area of existing realty units, regarding the temporal years when the realty was built in each city. This study creatively puts forward the overall prediction modeling and calculation framework of “grid cumulative prediction + incremental prediction correction for a city”. It increases the optimization options of grid granularity adjustment and local-global prediction correction for the artificial intelligence modeling and prediction of provincial realty big data and improves the applicability of the prediction model. Application analyses show that the calculation results of the modeling prediction system are reasonable and practical.

Abstract:To strengthen the management and control of Internet number resources such as IP addresses and autonomous system numbers (ASNs), the Internet engineering task force (IETF) proposes resource public key infrastructure (RPKI). In recent years, it has effectively solved the problems of route hijacking and path tampering and plays a crucial role in ensuring the stable operation of inter-domain routing. However, the security problems in the RPKI management mode are gradually highlighted, such as single point of failure, abnormal resource allocation, and verification failures caused by the poor synchronization of certificate revocation data. To tackle these problems, this study proposes a scheme for managing Internet number resources based on modifiable Blockchain. The experimental results show that the scheme is effective and feasible.

Abstract:Current research on unstructured data processing focuses on technological realization in experiments, while the discussions on the overall architecture and the practical application path in financial investment-research businesses are far from enough. Considering this, this study proposes to develop and design the intelligent investment-research platform through the combination of big data, natural language processing (NLP), and knowledge graph (KG) and implements the platform in real financial investment-research scenarios. With the data collecting, storage, and computation operated through the distributed system Hadoop, the platform integrates traditional text processing technologies and mainstream AI algorithms to form a deep semantic understanding capacity. As a result, the platform is capable of extracting financial text efficiently and storing the information in the form of KGs, and it can also provide intelligent analysis in financial investment research by further exploration and prediction. Taking the tests related to municipal bonds as samples, this research has proved the validity of this platform. The results show that the platform can automatically perform various functions with high precision through the whole process, thus promoting the working efficiency in financial investment research.

Abstract:This work mainly studies the problem of using multiple unmanned aerial vehicles (UAVs) to search for victims cooperatively in indoor scenes where the location information of victims relying on the global positioning system may be unreliable. To this end, this study proposes a multi-agent reinforcement learning (MARL) based solution which focuses on the path planning studies when the UAV team assists the rescue. Compared with the traditional solution, the proposed solution has advantages in large-scale indoor rescue scenes, such as deploying multiple rescue UAVs and rescuing multiple victims. At the same time, this solution also considers the charging problem of the UAVs to ensure that the power of the UAVs is always sufficient. Specifically, due to the continuous changes of the rescue scene depth parameters in the model, the proposed solution simulates the path planning as a decentralized partially observable Markov decision process (Dec-POMDP). To optimize the UAV control strategy, this study also trains a double deep Q-learning network (Double DQN). Finally, the Monte Carlo method is used to verify that this solution can effectively cooperate with multiple UAVs in a large-scale indoor environment and maximize the collection of the location information stored in the mobile phone used by the victim.

Abstract:In this study, an active and passive control method of intelligent cars is proposed. The STC89C51RC and K66 dual chips are used to control intelligent cars. With the Bluetooth communication technology, the car is controlled on a mobile phone APP. Meanwhile, automatic obstacle avoidance of the car is achieved with the ultrasonic ranging technology. In addition, an infrared detection sensor is added to facilitate the automatic tracking of the car. Beacon light homing is accomplished with the image recognition technology and the low-power MT9V032 camera. The experimental results show that the mobile car delivers a good tracking performance under proper lighting conditions. When the speed of the car is 20 cm/s, the obstacle avoidance accuracy reaches 99% and beacon lights within 7.85 m away from the car can be identified at a stable speed of 3.1 m/s.

Abstract:After the Ministry of Public Security implemented the new reforms to streamline administration and delegate power, improve regulation, and upgrade services, a vehicle inspection service platform was designed to solve the problems raised by the majority of car owners regarding the annual inspection of motor vehicles, such as the difficulty of vehicle inspection, high costs, congestion in inspection stations, and lack of time. This study focuses on the study and implementation of the business process and function structure of the vehicle inspection service platform. Specifically, interactions with vehicle owners can be achieved by means of mini programs, and vehicle inspection status can be identified through Internet of Things technology. Drivers and inspection station employees can participate through the mobile APP, and multiple parties such as vehicle owners, drivers, and inspection stations can be integrated through the service platform. The test results show that the system integration test and performance test results are good and can meet the actual application requirements. The online trial shows that the vehicle inspection service platform can to a certain extent solve the practical problems of vehicle owners’ difficulties in vehicle inspection and the uneven distribution of inspection stations, boasting bright application and promotion prospects.

Abstract:The Ceph storage system suffers from the metadata server performance bottleneck and low file reading efficiency during small file storage. In view of this, the study takes advantage of the inherent data relevance between small files to extract related features with a lightweight pattern matching algorithm. On this basis, small files are merged, as a result of which the rationality of file merging is improved. The small files in the same merged file are stored in the client cache to improve the cache read hit rate. Experiments have verified that the proposed scheme can effectively enhance the access efficiency of small files.

Abstract:Large-screen data visualization is an expression of data analysis results and an important part of data-enabled decision making. Given the long development cycle and high cost of large-screen data visualization software, an easy-to-use large-screen data visualization tool C317DataUI is developed on the basis of the Vue front-end framework and the Echarts visual components. The interface layout of the visual components is conducted by drag-and-drop operation, and data configuration management is carried out via the data connection panel of the components. The tool also provides some scene templates that can quickly deliver the expression of large-screen data visualization and thereby meet the needs of industrial users for data visualization with a low cost and a high efficiency.

Abstract:With the improvement of people’s living standards, air conditioning has become an essential part of work and life. The traditional PID control technology for air conditioning is still widely used, but it is exposed to problems such as inaccurate parameter control, sudden change, and hysteresis. The research on how to accurately control air-conditioning parameters and improve the performance of air-conditioning control systems has become popular in the field of air-conditioning control. In response to these problems, this study proposes an air-conditioning intelligent control mechanism based on fuzzy inference, which realizes the fuzzy control of air conditioning through several steps including parameter fuzzification, rule base construction, and fuzzy inference. The experiment proves the feasibility of this method and further verifies its advantages compared with PID in dealing with problems other than accuracy. The system realization part offers the concrete interactive process of air-conditioning intelligent control.

Abstract:Connecting rods of automobile engines are the main transmission parts, the quality of which directly affects the transmission performance of engines and ultimately influences the safety of vehicles. To realize the simultaneous detection of geometric parameters, bending, and twisting of connecting rods, the multi-parameter quality detection based on machine vision is proposed for them. Specifically, a vision system for multi-parameter quality detection is constructed for the connecting rods. The image preprocessing based on multi-threshold analysis and homomorphic filtering is studied to remove the shadow and enhance the contrast of the engine connecting rod images. The geometric features such as lines and circles are detected in these images based on the sub-pixel level analysis and Hough transform. Then, the least square method is used to fit the parameters of these geometric features with a further analysis of quality parameters. In this way, the multi-parameter quality detection for the connecting rods of automobile engines is realized. An application example has proved the effectiveness of the proposed method.

Abstract:As the number of branches connected to edge devices increases, the public and private network data that needs to be processed explodes, causing the edge devices to be overloaded and affecting the normal interaction of data. To solve these problems, this study analyzes the root cause of the problems and proposes a solution that integrates network equipment virtualization technology with multi-protocol label switching and border gateway protocol technology. For the feasibility verification of the solution, with the help of laboratory equipment, the environment required for the solution is built, and the deployment of the solution is completed. Then, the solution’s availability, data access control and isolation, distributed data processing and load sharing are tested, and it is compared with traditional methods in ten dimensions such as equipment redundancy, scalability, and management. The test and comparison results show that the solution can realize the distributed data processing and load sharing on edge devices, and it is superior to traditional methods as an effective VPN solution.

Abstract:Machining feature recognition is the key technology to realize the integration of CAD/CAPP/CAM. To tackle the robustness problem of the traditional recognition pattern of machining features based on symbolic reasoning, this study proposes an automatic recognition method of machining features based on deep learning of machining surface point cloud data. Utilizing the PointNet point cloud recognition framework, the study constructs a convolutional neural network (CNN) for automatic recognition of machining features of machining surface point cloud data. By the collection of the machining surface sets from CAD models and sampling of them to form point cloud data, a three-dimensional point cloud data library is constructed which is suitable for the learning of the network framework. A recognizer of machining features can be obtained by the CNN network training, able to automatically recognize 24 kinds of machining features, with the accuracy being higher than 99%. The method is simple, efficient, and insensitive to the point cloud data with noise and defects. Furthermore, it has good robustness and recognition effect for the damage of machining surfaces caused by feature intersection.

Abstract:With the maturity of object detection models, tracking-by-detection has become the mainstream of multi-object tracking research. Assisted by the almost perfect object detection results, data association can be formed only through the IoU information. However, in practice, a small number of missing detections will cause a large number of ID switches and fragmentations, which will seriously affect the tracking results. To solve this problem, the multiple object tracking algorithm is proposed with the introduction of image information. Specifically, preliminary tracking results obtained through the IoU model are verified with the pedestrian feature vector, and for the tracks that have not passed the verification, they are re-matched. For the problem of occlusion, the algorithm adopts the method of predicting the object trajectory and taking measures in advance. Tested on MOT16 and 2DMOT15 datasets, the algorithm has achieved good results, and it is more suitable for practical applications with its online tracking mode.

Abstract:In view of the sparse reward problem in the training of energy consumption control systems using reinforcement learning methods, a deep deterministic policy gradient (DDPG) method based on the self-supervised network is applied to the building energy consumption control. First, the processing state and action variables are regarded as the input of the self-supervised network forward model, predicting the feature vector of the next state and using the prediction error as the internal reward of curiosity to solve the sparse reward problem. Then, a data-driven method is used to train the building energy consumption model with weather data as input and energy consumption data as output. Finally, the DDPG method based on the self-supervised network is used to develop the optimal control strategy, and the optimal discharge temperature of the air handling unit (AHU) is set based on the strategy to reduce the energy consumption of the equipment. Experimental results show that this method can achieve good energy-saving effects on the basis of maintaining a comfortable building environment.

Abstract:This study applies deep reinforcement learning to the nesting problem of two-dimensional irregular polygons. The shape characteristics of polygons are mapped into one-dimensional vectors according to the distances from the centroid to the contours. For randomly generated polygons, the compression losses are less than 1%. With a given sequence of the polygon items, this study employs a multi-task deep reinforcement learning model to predict the sequence and rotation angle of the irregular nesting items and obtains a nesting result 5%–10% higher than those of the traditional heuristic algorithms. A result better than that of the optimized genetic algorithm is also achieved under a sufficient sampling number. The model can deliver a better initial solution in the shortest time and, therefore, has a generalization ability.

Abstract:To avoid the danger of people using mobile phones while walking, this study proposes a lightweight model (Mobile-YOLOv3) with strong real-time performance to detect road obstacles. We photograph roadblocks and annotate a roadblock data set around Guangzhou City. Lightweight is achieved by the replacement of the backbone network of YOLOv3 with a lightweight MobileNetv1 network. In addition, we apply four methods to improve detection accuracy and model robustness, i.e., border regression loss function CIOU, classification loss function Focal, prediction box screening algorithm Soft-NMS, and negative sample training. The experimental results show that the model obtains 98.84% MAP. Compared with YOLOv3, this model has the scale reduced by 2.5 times but the detection accuracy improved by 7%.

Abstract:In the generation of data classification prediction models, highly unbalanced training data will significantly degrade the performance of the model. Therefore, this study proposes an improved oversampling method for unbalanced data sets based on genetic ideas. Inspired by the chromosome theory of inheritance in biology, this method uses close relatives to generate similar but not identical new instances to balance the majority of classes. Under the premise of the same sample distribution, the bias influence of unbalanced data on the training results is reduced or even eliminated. Finally, a comparative experiment on a public data set shows that the method has achieved a higher recall rate and G-mean value, which proves that the improved method is effective and the comprehensive performance of the generated model has been promoted.

Abstract:To address the problems of the existing deep-learning defogging algorithm such as the various parameters, long training time, and inability to apply to real-time computer vision systems, this study proposes a bright and dark channel CycleGAN network (BDCCN). BDCCN, based on the CycleGAN, improves the cyclic perceptual loss and achieves image defogging by combining the fixed parameters with training parameters and drawing on the priori theory of bright and dark channels. The experimental results show that the algorithm proposed in this paper, with a small amount of calculation and a fast convergence rate, performs well on both synthetic data sets and real data sets.

Abstract:The traveling salesman problem (TSP) is a classical NP-hard problem. The research on it has never stopped, and a lot of approximate solving algorithms have been obtained. However, each algorithm has its own characteristics, and thus new algorithms are proposed frequently for TSP, such as the sparrow algorithm developed recently. This work studies and analyzes the principle, search strategy, and basic process of the sparrow search algorithm (SSA). When the search by SSA approaches the global optimum, the diversity of the population decreases and it is easy to fall into the local optimum. Given this, the work proposes an improved sparrow search algorithm (ISSA). Six standard test functions, the basic SSA, and other swarm intelligence algorithms are employed in simulation experiments to test the performance of ISSA. Finally, ISSA is used to solve the TSP. Experiments show the effectiveness of ISSA in improving the shortcomings of SSA and enhancing the optimization ability and verify the feasibility and superiority of ISSA in TSP solving.

Abstract:Pose estimation has always been a key issue in the field of 3D reconstruction. A tightly coupled real-time pose optimization method for the mobile terminal is proposed to ensure the real-time performance under the limited resources of the mobile terminal and improve the accuracy of trajectory calculation. First, image information and motion sensor information are obtained to conduct pretreatments such as feature extraction and pre-integration. Then, the reprojection error and the inertial sensor error are calculated according to the epipolar geometric constraints. Finally, the weighted error is used to jointly optimize the calculation of the pose trajectory. The tight coupling strategy can efficiently use the consistency in the pose constraint of image information and inertial motion information. Experiments on the public data set EuRoC show that compared with the existing visual-inertial pose estimation methods, the proposed method guarantees the real-time performance on the mobile terminal and has a smaller camera trajectory error in reconstruction.

Abstract:At present, PCB welding defect images screened with traditional machine vision analysis methods still need manual reinspection, which is easy to make mistakes after visual fatigue due to heavy workload. In view of this, the study designs and applies the YOLOv3-spp object detection algorithm to build a welding defect detection model. For a higher detection speed, model pruning, model distillation, model quantization and other technologies are used to compress and optimize the detection model. OpenVINO, a deep learning acceleration component, is employed to load the compressed and optimized detection model for the reinspection of PCB welding defect images. With the help of this optimization algorithm, this study designs a PCB welding defect detection and identification system based on deep learning technology. It can quickly and accurately identify welding defects and locate the defects, addressing the low efficiency and high rates of missed detection and false detection caused by manual visual inspection.

Abstract:With the development of computer performance, pre-trained machine learning models are used for inference on personal devices. Caffe is a popular deep learning framework featuring image classification. However, it can only infer using one CPU core or one GPU if without customization, which limits the computing power of heterogeneous parallel computation devices. Deep learning is a demanding task for a computation device. For a better user experience and faster inference, it is important to fully use all computing cores of the device via parallelization. Considering the CPU-to-GPU performance ratio may vary on different deep learning models, tasks should not just be equally assigned to all computing cores. It should be noted that more overhead will be introduced if the tasks are divided into too many portions or synchronized scheduling algorithms are used. Thus, a well-designed scheduling algorithm able to reduce idle time is crucial for better performance. Some approaches have been developed to improve Caffe performance on heterogeneous parallel computation devices, whereas there are some limits on the platform hardware and usage. As a result, it is difficult to fully utilize the performance of these devices. This study reports the work on the improvement of Caffe interface and the proposed new algorithms. Caffe interface is extended to enable customized programs to use multiple computing cores or devices of a heterogeneous parallel platform for deep learning inference with Caffe. Some existing scheduling algorithms are ported and tested. To avoid synchronization overhead, two novel asynchronous scheduling algorithms, async-FIFO and fast-split, are proposed. All scheduling algorithms are tested and results show that the Caffe inference performance of heterogeneous parallel computation devices adopting fast-split is significantly faster than that in the case where only one computing core is adopted. Fast-split on average reduces performance waste by 7.4% and 21.0% on MNIST and Cifar-10 datasets, respectively, compared with the current best heterogeneous parallel scheduling algorithm HAT.

Abstract:The current assisted pneumonia diagnosis method using the residual network (ResNet) based on batch normalization has high dependence on the batch size and a low utilization rate of network channel features, and pneumonia diagnosis methods using deep neural networks all ignore the problems of medical data privacy and islands. To solve these problems, this study proposes an assisted diagnosis method that integrates the federated learning framework, the squeeze-and-excitation network, and the improved ResNet (FL-SE-ResNet-GN). This method uses FL to protect data privacy and pays full attention to channel characteristics with the SE network and the group normalization method. Experimental results on the Chest X-Ray Images dataset show that the accuracy, precision, and recall of this method reach 0.952, 0.933, and 0.974, respectively. Compared with other existing methods, this method has significantly improved the accuracy and recall indicators on the basis of protecting data privacy.

Abstract:License plate image reconstruction plays an important role in the intelligent transportation system. After repeated experiments, a super-resolution image reconstruction method for license plates is proposed with the help of generative adversarial networks (GANs). The method mainly consists of four parts: (1) pretreatment of the input image, including image resizing and filtering of images with poor contrast; (2) image feature extraction using a residual dense network; (3) introduction of progressive sampling, which can provide a larger receptive field and more information details; (4) introduction of a discriminator based on PatchGAN to make a more accurate judgment, which guides the generator to reconstruct images with higher quality and more details. The comparison with a current superior algorithm on the Chinese City Parking Dataset (CCPD) proves that the proposed model has higher PSNR and SSIM (26.80 and 0.77, respectively) and less time of reconstructing a single-frame image (only 0.06 s), which verifies the feasibility of the proposed approach in license plate image reconstruction.

Abstract:To address the prediction problems of reservoir grain sizes in low permeability oilfields, this study proposes a scheme for predicting reservoir grain sizes in low permeability oilfields with the extreme gradient boosting (XGBoost) in machine learning. First, a proper XGBoost model is built in consideration of the problems. Then, well logging curves suitable for grain size prediction are selected to create a sample database according to the established relationships of the characteristic values of the core reservoir grain size with other logging information. Finally, sample database data are employed to train the newly built XGBoost model. The trained model can predict unknown reservoir grain size characteristics in a study area. The results show that the XGBoost model designed in this study is superior to the back propagation (BP) neural network in calculation efficiency and prediction accuracy of reservoir grain sizes in low permeability oilfields.

Abstract:With the real border gateway protocol (BGP) update message data disclosed on the Internet, this study proposes a new BGP anomaly detection method based on graph embedding features and long short-term memory (LSTM) AutoEncoder, which focuses on the network topology and variation characteristics in time series. First, the AS_PATH attribute information of BGP data is used to construct a dynamic embedding feature dataset based on the network topology of time series, and then the LSTM AutoEncoder model is employed for data detection to find abnormal ones. For the actual data of abnormal events, the method successfully detects the abnormal data and has higher accuracy than traditional detection methods.

Abstract:Chinese calligraphy is one of the representatives of Chinese traditional culture. However, the different styles, complex structures, and various distortions of calligraphic fonts have brought great obstacles to learning and appreciating calligraphy for the public. A calligraphic font recognition algorithm based on an improved DenseNet network is proposed to solve the difficulty of ordinary people in interpreting calligraphy works. A regional weight ratio pooling rule is designed to replace the maximum pooling and average pooling rules of the traditional DenseNet network. The Nadam algorithm is used to adjust the adaptive learning rate and optimize the model training effect. In addition, a model pruning strategy based on the pruning technology is proposed, which ensures a strong recognition performance and improves the training efficiency of the model. The experimental results show that in a mixed font data set composed of four types of fonts, namely the standard script, the running script, the clerical script, and the seal script, the proposed algorithm obtains a recognition rate of 96.13%, which is better than those of the other five deep learning models.

Abstract:Given that single-cycle frequency detection cannot make full use of the cyclic spectrum information in spectrum sensing, this study proposes a multi-cycle frequency cooperative spectrum sensing method based on radio environment map (REM) information. The first step of this method is to select multiple same cyclic frequencies at multiple cognitive users for cyclostationary detection. The spectral correlation function amplitude is adopted as the detection statistic. The threshold value at a constant false-alarm rate (CFAR) is set according to the derived decision threshold formula, and the detection result at a single cognitive user is obtained through decision fusion. In the second step, the weight coefficient at each cognitive node is calculated according to the distance between the authorized user and the cognitive user provided by the REM. Weighted fusion of the weight coefficients and the detection results of the corresponding nodes are conducted to improve the reliability of the detection results. Simulation results show that the improved method can effectively detect authorized users and has better detection performance and stronger practicability under a low signal-to-noise ratio (SNR).

Abstract:In the task of text classification, traditional natural language processing methods have limitations in short text classification due to the sparse features and irregular wording of short texts. Considering the characteristics of short texts, this study proposes a classification algorithm based on the fusion of bidirectional encoder representations from Transformers (BERT) and a collapsed Gibbs sampling algorithm for the Dirichlet multinomial mixture model (GSDMM) and clustering guidance to improve the effectiveness and accuracy of short text classification. First, the model converts short texts into integrated semantic vectors by using the fusion model of BERT and GSDMM. The integrated vectors reflect global semantic features and topic features and solve the problems of sparse short text features and the lack of topic information. Then, the clustering guidance algorithm is introduced into the front-end training of the classifier, which realizes the expansion of the labeled data and improves the interpretability of the results. Finally, the expanded labeled data set is used to train the classifier to complete the automatic classification of short texts. Taking the negative comment of an e-commerce platform as the verification data set, this study verifies the effectiveness and advantages of the algorithm in short text classification in multiple groups of comparative experiments.

Abstract:China is a large agricultural country. In the process of agricultural production, it is of great significance to accurately predict the soil moisture. In view of the local minimization and slow convergence in the prediction process of the traditional back propagation (BP) neural network, an improved genetic algorithm is applied to the traditional BP neural network model in this study. A soil moisture prediction method is proposed that optimizes the BP neural network by the adaptive genetic algorithm. A prediction model of the BP neural network optimized by the improved genetic algorithm is established by the Matlab simulation software and experimented on the soil moisture of corn fields in Harbin. The results show that the accuracy of the model is higher than that of the unoptimized BP neural network model. This model can greatly reduce the use of moisture sensor and thus reduce the agricultural production cost.

Abstract:With the development of economy, robotic of economy, robotic cells have greatly improved the production efficiency and quality of the manufacturing industry. Compared with that of the traditional flexible manufacturing cells, the scheduling problem of shops with robot handling also involves material handling. As a result, the production scheduling problem is becoming increasingly complex. In view of the lack of dominance of the Pareto dominance relation in high-dimensional multi-objective optimization, the Lorenz domination and CDAS domination are combined, respectively, with the non-dominated sorting genetic algorithm-III (NSGA-III) algorithm in this study and applied for the first time to the high-dimensional multi-objective scheduling of shops with robotic cells. Considering the complexity of modern production processes, this study proposes to optimize multiple objectives, such as maximum completion time, total processing energy consumption, delivery lead time, delivery delay time, and total production cost, at the same time to determine the operating state and handling sequence of the robot and improve production efficiency. Experiments show that on the abovementioned production scheduling problem, the NSGA-III algorithm based on Lorenz domination or CDAS domination performs better than the traditional NSGA-III algorithm in the solution convergence and uniformity.

Abstract:In view of the lack of security assessment methods of LTE-R communication systems in high-speed railway and the fuzziness and gray of the evaluation process, this study develops a risk assessment model for LTE-R communication systems based on set pair extension and improved analytic hierarchy process. The evaluation system is established mainly from the aspects of people, equipment, and networks that affect the communication system, and the risk assessment indicators and evaluation standards are determined. The improved analytic hierarchy process is used to calculate the weight, construct the system risk set pair via the set pair extension and calculate the risk connection membership degree to judge the risk level of the communication system. The analysis of an example shows that the use of set pair extension and improved analytic hierarchy process can effectively and truly reflect the risk state of LTE-R communication systems and can provide a theoretical basis for their security construction and risk control.

Abstract:Long text matching is a basic work of natural language processing, and it plays a key role in text clustering, news recommendation, etc. Due to the limitations of the corpus, space structure, and text representation technology, long text matching has been progressing slowly. The bidirectional encoder representations from Transformer (BERT) model proposed in recent years has an excellent performance in the text representation. For BERT, there are three common methods for processing long texts: truncation, segmentation, and compression. The truncation method causes the loss of massive text information; the segmentation method retains text information but loses part of the semantic information; the compression method may lose part of the key information. In response to the above problems, this study improves the segmentation method and proposes a long text matching model based on BERT (LTM-B), which is based on the Siamese neural network and adopts a layered idea to divide the document into multiple segments. The BERT model is used for text vectorization. As a result, the matrix representation of the document is obtained. The bidirectional long short-term memory (BiLSTM) is employed to generate the position matrix, and then the sum of the document matrix and the position matrix is input to the Transformer encoder for feature extraction. Finally, the two matrices are interacted, pooled, and spliced, and then the matching results are output through the fully connected layer classification. Experiments show that the LTM-B model outperforms other methods in long text matching.

Abstract:Emotion recognition is closely related to many facets of our daily lives. However, it is difficult to achieve a satisfying emotion recognition rate by using one single algorithm. Therefore, this study puts forward an emotion recognition model based on electroencephalogram (EEG) with a fusion algorithm that combines the support vector machine (SVM) algorithm with the K-nearest neighbors algorithm (SVM-KNN). In the emotion classification process, the spatial distance between the sample to be identified and the optimal classification hyperplane is calculated. If it is longer than the preset threshold, the SVM classifier is chosen to classify the emotion records. Otherwise, the KNN classifier is chosen. Finally, experiments are carried out on the SJTU emotion EEG dataset (SEED). The comparative experiments show that the SVM-KNN algorithm improves the accuracy of the three-emotion classification. This model can effectively identify the types of emotions and thus has positive significance in obtaining the emotions of patients with expression disorders in medical care.

Abstract:To tackle the problem that density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm is increasingly time-consuming with the increase in data volume, this study proposes an improved DBSCAN algorithm based on a K-dimensional (KD) tree (hereinafter referred to as KD-DBSCAN). The KD tree is used to divide the data set, construct the neighborhood object set, and distinguish the noise point and the core point in advance to avoid the calculation of the noise neighborhood set in the clustering process and speed up the neighborhood set query of the core point object. In this study, the global positioning system (GPS) data of a floating car is used as experimental data to compare the traditional DBSCAN algorithm and KD-DBSCAN algorithm in aspects of the clustering effect and time performance. The experimental results show that the KD-DBSCAN algorithm is comparable to the traditional DBSCAN algorithm in the clustering effect but has greatly improved time performance.

Abstract:In the optimization design and analysis of expandable casing joints, it is necessary to modify the joint parameters repeatedly and establish the 3D model of the joint efficiently and quickly. This paper presents two parametric modeling methods of expandable casing joints with Unigraphics (UG). After the joint model template is completed in UG, one of the two methods uses the development platform Visual Studio to write an executable program with modifiable structure parameters and thereby achieve the secondary development and modeling of the expandable casing joint. In the other method, a dialog box is designed on the basis of the PTS module and parametric modeling is performed through the reuse library. An example of 3D parametric modeling of expansion casing joints is also provided in this paper. The proposed parametric modeling method can effectively reduce the repeated modeling time in the numerical analysis of expandable casing joints and improve work efficiency.

Abstract:In the BGP protocol plaintext transmission, attackers easily forge the prefix and path information, which thereby causes prefix hijacking with great harm. The AS path information protection mainly involves two aspects: path tamper-proofing and verification of illegal content. Resource public key infrastructure (RPKI) is an important security system to solve route hijacking. Currently, the path verification solutions under the RPKI system mainly include BGPSec, ASPA and Path-End, among which BGPSec mainly addresses path tampering, while ASPA and Path-End target path legality verification. However, these schemes have the defects of complicated calculation or weak path protection. A small number of signatures are introduced into the ASPA scheme to improve the granularity limiting path tampering. Therefore, this study proposes an improved path protection mechanism and designs comparison experiments with other schemes regarding the overhead and safety performance. The experimental results show that the performance of the improved scheme is better than that of the other schemes under the condition of introducing limited overhead.

Abstract:In the reinforcement learning method for the continuous integration test case prioritization (CITCP), the agent rewards the test cases to realize the adjustment of test case prioritization strategy, and thus they can meet the needs of frequent iteration and rapid feedback in continuous integration testing. The agent usually only rewards the failure test cases. However, in the actual industrial processes, the continuous integration testing features high-frequency integration and low-failure-rate tests, which poses a new challenge to the actual application of CITCP. Low-failure-rate tests can be understood as a sparse number of failure test cases, which can lead to the sparsity of reward objects in reinforcement learning and bring about the sparse reward problem. In this study, a reward object selection strategy is proposed to solve the sparse reward problem. With the failure test cases rewarded, passing test cases similar to failure test cases are selected to be rewarded, and thus the number of reward objects increases. Specifically, the similarity measure method for test cases is designed with the feature vector representation of historical execution information sequences and duration time. Then, the passing test cases similar to the failure test cases are selected to be rewarded through the similarity measure. The experiments are conducted in six industrial data sets, and the results show that the similarity-based reward object selection strategy can effectively solve the sparse reward problem by increasing the reward objects and further improve the quality of reinforcement learning-based CITCP.

Abstract:Thanks to the augmented reality technology, virtual information can be embodied and integrated in the real world, enabling the increasingly wide applications of virtual information detached from the physical world in real-world scenarios. On this basis, this study proposes an efficient real-time virtual try-on technique that can be applied to a variety of practical scenarios. For example, in the e-commerce scenario, users can select the corresponding model file of a certain style online for virtual try-on before purchasing the product and make their decisions upon the virtual try-on results. The proposed method maps the model file to a graphical state that can be added with the real-time video stream on the basis of the face pose parameters. After the addition is performed in a specific region, the result is fed back to the video frame. The final added model file is able to adapt itself to the position change of the head. The experimental results show that the proposed method delivers a good performance in face distance and position, graphics rendering, and real-time wearability.

Abstract:The segmentation of eyeball areas is a key step in medical ultrasound image processing and analysis. Since the eyeball ultrasound images collected by clinical equipment have disadvantages including noise interference, blurred areas, and similar edge gray levels, the existing methods cannot accurately segment eyeball areas. Therefore, this study proposes a semantic embedded attention mechanism for eyeball segmentation based on deformable convolutions. Firstly, deformable convolutions, instead of traditional convolutions, are used to improve the representational ability of the network in eyeball areas. Secondly, a semantic embedded attention mechanism is constructed to fuse semantic information among different layers, enhance the salient features in the target area, and reduce the wrong segmentation of the background area, thereby improving the segmentation accuracy of the network. Finally, in order to check the segmentation performance, the proposed model in this study is compared with three existing deep learning segmentation models, and it obtains the highest accuracy on the segmentation data set of ultrasound eyeball images, fully verifying that this model has better segmentation ability and robustness.

Abstract:To address the low throughput, high latency and random selection of master nodes in the Blockchain-based fruit quality traceability system, this study proposes an improved practical Byzantine fault tolerance (PBFT) consensus algorithm based on integral selection. The algorithm introduces the integral selection protocol and optimizes the consistency protocol, view change protocol and garbage collection mechanism to improve the probability of honest master nodes being selected and reduce the communication overhead between nodes, thus improving the efficiency of consensus algorithm execution. At the same time, when the garbage collection mechanism is operated, the integrals are reallocated to all participating nodes for the dynamic change in the node number. Experiments show that the method proposed in this study has better performance in improving the throughput and reducing the latency of consensus algorithms.

Abstract:With the cost reduction and popularization of digital cameras, non-contact structural health monitoring technology based on computer vision has been paid more and more attention to. Traditional contact measuring instruments may cause mass loads on lightweight structures, and the installation and maintenance on large civil structures are costly and time-consuming, especially for long-term applications. As an alternative non-contact method, the computer vision method using digital cameras has relatively low cost and is flexible. In order to solve the problem of low accuracy and stability of the traditional optical flow method in micro displacement measurement, this study proposes an improved phase-based optical flow method to calculate the vibration displacement of structures. In this method, the partial complex steerable filter constructed in the phase-based video motion magnification method is introduced to convolute with the video sequence image to obtain the local spatial phase information of structures. Then, the phase-based optical flow method is used to calculate the pixel displacement of structural vibration. Finally, the pixel displacement is transformed into real displacement by noise reduction and the scale factor method. The experimental results show that the proposed method has higher accuracy and stability than other optical flow methods.

Abstract:As the resource public key infrastructure (RPKI) coverage of the inter-domain network expands, the consistency of RPKI data synchronization in the actual deployment, the risk of operational errors and abuse of authority power have become major obstacles to the full deployment of RPKI. This study presents a scheme for detecting conflicts of updating RPKI cache based on fact ownership of route origin. This scheme uses reverse RTR protocol and multi-layer transmission architecture of RPKI data to collect and synchronize fact route origin information. Then, it compares fact route origin information and RPKI cache update data to detect conflicting data of RPKI cache update, which ensures authenticity and effectiveness of RPKI cache. Finally, the data synchronization efficiency and detection performance of this scheme are compared with those of other schemes. The experimental results show that this scheme has some detection advantages.

Abstract:The intelligent recognition of graphite is particularly critical to the transformation of the mining industry to informatization and intellectualization. To address the long time and low efficiency in manually identifying graphite, this study proposes an improved AlexNet network for graphite image recognition. First, image preprocessing is performed on the data set through random cropping, horizontal flipping according to probability, and normalization to achieve data augmentation. Then, the activation function ReLU6 is employed to compress the dynamic range so that the algorithm can become more robust. The batch standardization algorithm is used for normalization to speed up the convergence, and the convolution kernel is resized to enhance the generalization ability. Finally, dropout regularization is added to the fully connected layer to further prevent overfitting. Compared with the existing method, the proposed method reduces the loss value and improves the average accuracy of graphite recognition in the simulation experiment.