Most Download articles

    Published in last 1 year | In last 2 years| In last 3 years| All| Most Downloaded in Recent Month | Most Downloaded in Recent Year|

    Most Downloaded in Recent Year
    Please wait a minute...
    For Selected: Toggle Thumbnails
    Review of Application of Transfer Learning in Medical Image Field
    GAO Shuang, XU Qiaozhi
    Computer Engineering and Applications    2021, 57 (24): 39-50.   DOI: 10.3778/j.issn.1002-8331.2107-0300
    Abstract288)      PDF(pc) (896KB)(432)       Save

    Deep learning technology has developed rapidly and achieved significant results in the field of medical image treatment. However, due to the small number of medical image samples and difficult annotation, the effect of deep learning is far from reaching the expectation. In recent years, using transfer learning method to alleviate the problem of insufficient medical image samples and improve the effect of deep learning technology in the field of medical image has become one of the research hotspots. This paper first introduces the basic concepts, types, common strategies and models of transfer learning methods, then combs and summarizes the representative related research in the field of medical images according to the types of transfer learning methods, and finally summarizes and prospects the future development of this field.

    Reference | Related Articles | Metrics
    Attention-YOLO:YOLO Detection Algorithm That Introduces Attention Mechanism
    XU Chengji, WANG Xiaofeng, YANG Yadong
    Computer Engineering and Applications    2019, 55 (6): 13-23.   DOI: 10.3778/j.issn.1002-8331.1812-0010
    Abstract2089)      PDF(pc) (1426KB)(4474)       Save
    YOLOv3 is a real-time object detection algorithm, its speed and accuracy reach good trade-off, but the disadvantages are that the boundary box positioning is inaccurate and it is difficult to distinguish overlapping objects. For the above problems, this paper proposes the Attention-YOLO algorithm based on the item-wise attention mechanism which embeds channel and spatial attention mechanism in the feature extraction network, uses the filtered weighted feature vector to replace the original residual fusion, and adds a second-order item to reduce the information loss in the process of fusion and accelerate the convergence of the model. Based on the experiments on COCO and PASCAL VOC datasets, the results show that the Attention-YOLO algorithm effectively reduces the boundary box positioning loss and improves the detection accuracy. Compared with YOLOv3, the Attention-YOLO improves at most 2.5 mAP@IoU[0.5∶0.95] on COCO dataset, and reaches 81.9 mAP on PASCAL VOC 2007 test.
    Related Articles | Metrics
    Research Progress of Transformer Based on Computer Vision
    LIU Wenting, LU Xinming
    Computer Engineering and Applications    2022, 58 (6): 1-16.   DOI: 10.3778/j.issn.1002-8331.2106-0442
    Abstract355)      PDF(pc) (1089KB)(338)       Save
    Transformer is a deep neural network based on the self-attention mechanism and parallel processing data. In recent years, Transformer-based models have emerged as an important area of research for computer vision tasks. Aiming at the current blanks in domestic review articles based on Transformer, this paper covers its application in computer vision. This paper reviews the basic principles of the Transformer model, mainly focuses on the application of seven visual tasks such as image classification, object detection and segmentation, and analyzes Transformer-based models with significant effects. Finally, this paper summarizes the challenges and future development trends of the Transformer model in computer vision.
    Reference | Related Articles | Metrics
    Review of Development and Application of Artificial Neural Network Models
    ZHANG Chi, GUO Yuan, LI Ming
    Computer Engineering and Applications    2021, 57 (11): 57-69.   DOI: 10.3778/j.issn.1002-8331.2102-0256
    Abstract536)      PDF(pc) (781KB)(664)       Save

    Artificial neural networks are increasingly closely related to other subject areas. People solve problems in various fields by exploring and improving the layer structure of artificial neural networks. Based on the analysis of artificial neural networks related literature, this paper summarizes the history of artificial neural network growth and presents relevant principles of artificial neural networks based on the development of neural networks, including multilayer perceptron, back-propagation algorithm, convolutional neural network and recurrent neural network, explains the classic convolutional neural network model in the development of the convolutional neural network and the widely used variant network structure in the recurrent neural network, reviews the application of each artificial neural network algorithm in related fields, summarizes the possible direction of development of the artificial neural network.

    Related Articles | Metrics
    Review of Application Prospect of Deep Reinforcement Learning in Intelligent Manufacturing
    KONG Songtao, LIU Chichi, SHI Yong, XIE Yi, WANG Kun
    Computer Engineering and Applications    2021, 57 (2): 49-59.   DOI: 10.3778/j.issn.1002-8331.2008-0431
    Abstract601)      PDF(pc) (982KB)(1030)       Save

    As the latest development of machine learning, deep reinforcement learning has been shown in many application fields. The algorithm research and application research of deep reinforcement learning have produced many classical algorithms and typical application fields. The application of deep reinforcement learning in industrial manufacturing can realize high level control in complex environment. First of all, the research on deep reinforcement learning is summarized, and the basic principles of deep reinforcement learning are introduced, including deep learning and reinforcement learning. Then, the paper introduces the theoretical methods of the application of deep reinforcement learning algorithm. On this basis, it classifies the algorithms of deep reinforcement learning, respectively introduces the reinforcement learning algorithm based on value function and the reinforcement learning algorithm based on strategy gradient, and lists the main development results of these two kinds of algorithms, as well as other related research results. Then, the typical applications of deep reinforcement learning in industrial manufacturing are classified and analyzed. Finally, the existing problems and future development direction of deep reinforcement learning are discussed.

    Related Articles | Metrics
    Computer Engineering and Applications    2022, 58 (9): 0-0.  
    Abstract55)      PDF(pc) (38025KB)(263)       Save
    Related Articles | Metrics
    Robot Dynamic Path Planning Based on Improved A* and DWA Algorithm
    LIU Jianjuan, XUE Liqi, ZHANG Huijuan, LIU Zhongpu
    Computer Engineering and Applications    2021, 57 (15): 73-81.   DOI: 10.3778/j.issn.1002-8331.2103-0525
    Abstract242)      PDF(pc) (1452KB)(396)       Save

    Traditional A* algorithm is one of the commonly used algorithms for global path planning of mobile robot, but the algorithm has low search efficiency, many turning points in planning path, and can’t achieve dynamic path planning in the face of random dynamic obstacles in complex environment. To solve these problems, the improved A* algorithm and DWA algorithm are integrated on the basis of global optimization. The obstacle information in the environment is quantified, and the weight of heuristic function of A* algorithm is adjusted according to the information to improve the efficiency and flexibility of the algorithm. Based on the Floyd algorithm, the optimization algorithm of path nodes is designed, which can delete redundant nodes, reduce turning points and improve the path smoothness. The dynamic window evaluation function of DWA algorithm is designed based on the global optimal, which is used to distinguish known obstacles from unknown dynamic and static obstacles, and the key points of the improved A* algorithm planning path are extracted as the temporary target points of DWA algorithm. On the basis of the global optimal, the fusion of the improved A* algorithm and DWA algorithm is realized. The experimental results show that, in the complex environment, the fusion algorithm can not only ensure the global optimal path planning, but also effectively avoid the dynamic and static obstacles in the environment, and realize the dynamic path planning in the complex environment.

    Related Articles | Metrics
    Overview on Video Abnormal Behavior Detection of GAN via Human Density
    SHEN Xulin, LI Chaobo, LI Hongjun
    Computer Engineering and Applications    2022, 58 (7): 21-30.   DOI: 10.3778/j.issn.1002-8331.2110-0364
    Abstract98)      PDF(pc) (1720KB)(245)       Save
    As an important branch of computer vision, video anomaly detection is a challenging task for intelligent video surveillance systems. It is generally referred to as automatic recognition of videos that contain abnormal targets, events or behaviors, which plays a vital role in ensuring public safety. Generative adversarial network(GAN) is anemerging unsupervised method, which can not only be used to generate images, its unique adversarial learning idea also shows good development potential in the field of anomaly detection. Firstly, the framework of the GAN is introduced. Secondly, according to the density of the scene and the object on which the action is taking place, the research status of video anomaly detection based on GAN is discussed from two aspects of individual behavior anomalies, group anomalies. These two types of abnormalities are further elaborated on the basic of reconstruction and prediction methods respectively. Thirdly, the common datasets for video anomaly detection are briefly introduced, finally, the future development is prospected.
    Reference | Related Articles | Metrics
    Survey of Interpretability Research on Deep Learning Models
    ZENG Chunyan, YAN Kang, WANG Zhifeng, YU Yan, JI Chunmei
    Computer Engineering and Applications    2021, 57 (8): 1-9.   DOI: 10.3778/j.issn.1002-8331.2012-0357
    Abstract379)      PDF(pc) (677KB)(582)       Save

    With the characteristics of data-driven learning, deep learning technology has made great achievements in the fields of natural language processing, image processing, and speech recognition. However, due to the deep learning model featured by deep networks, many parameters, high complexity and other characteristics, the decisions and intermediate processes made by the model are difficult for humans to understand. Therefore, exploring the interpretability of deep learning has become a new topic in the current artificial intelligence field. This review takes the interpretability of deep learning models as the research object and summarizes its progress. Firstly, the main interpretability methods are summarized and analyzed from four aspects:self-explanatory model, model-specific explanation, model-agnostic explanation, and causal interpretability. At the same time, it enumerates the application of interpretability related technologies, and finally discusses the existing problems of current interpretability research to promote the further development of the deep learning interpretability research framework.

    Related Articles | Metrics
    Algorithm for Portrait Segmentation Combined with MobileNetv2 and Attention Mechanism
    WANG Xin, WANG Meili, BIAN Dangwei
    Computer Engineering and Applications    2022, 58 (7): 220-228.   DOI: 10.3778/j.issn.1002-8331.2106-0334
    Abstract90)      PDF(pc) (1064KB)(235)       Save
    As for low precision and efficiency in portrait segmentation, an algorithm for portrait segmentation combined with MobileNetv2 and attention mechanism is proposed to achieve the portrait segmentation. With keeping the encoder-decoder of U-typed network , MobileNetv2 is used as the backbone of the network and streamline the upsampling process, it can reduce the parameters of the network. It is helpful for transfer and network training. The network with attention mechanism can learn portrait features more effectively, and the mixed loss is beneficial to the classification of difficult pixels of portrait edges. A portrait bust can be selected as the input of the model, and the corresponding image mask can be produced by the network. The proposed algorithm is tested on Human_Matting dataset and EG1800 dataset. The results show that the accuracy of the proposed algorithm is 98.3%(Matting) and 97.8%(EG1800), which is higher than PortraitNet(96.3%(Matting) and 95.8%(EG1800)) and DeepLabv3+(96.8%(Matting) and 96.4%(EG1800)). The algorithm can clearly separate the target person from the background. The proposed algorithm’s IOU can reach to 98.6%(Matting) and 98.2%(EG1800), which can be used in lightweight applications and provides a new research idea for portrait segmentation.
    Reference | Related Articles | Metrics
    Review of Cognitive and Joint Anti-Interference Communication in Unmanned System
    WANG Guisheng, DONG Shufu, HUANG Guoce
    Computer Engineering and Applications    2022, 58 (8): 1-11.   DOI: 10.3778/j.issn.1002-8331.2109-0334
    Abstract226)      PDF(pc) (913KB)(232)       Save
    As the electromagnetic environment becomes more and more complex as well as the confrontation becomes more and more intense, it puts forward higher requirements for the reliability of information transmission of unmanned systems whereas the traditional cognitive communication mode is difficult to adapt to the independent and distributed development trend of broadband joint anti-interference in future. For the need of low anti-interference intercepted communications surrounded in unmanned systems, this paper analyzes the cognitive anti-interference technologies about interference detection and identification, transformation analysis and suppression in multiple domains and so on. The research status of common detection and estimation, classification and recognition are summarized. Then, typical interference types are modeled correspondingly, and transformation methods and processing problems are concluded. Furthermore, traditional interference suppression methods and new interference suppression methods are systematically summarized. Finally, the key problems of restricting the joint interference of broadband are addressed, such as the classification and recognition of unknown interference, the temporal elimination of multiple interference, the joint separation of distributed interference and the optimal control of collaborative interference, which highlight the important role of cognitive interference suppression technology in unmanned system communication.
    Reference | Related Articles | Metrics
    Summary of road extraction methods for remote sensing images
    ZHANG Yonghong, HE Jing, KAN Xi, XIA Guanghao, ZHU Linglong, GE Taotao
    Computer Engineering and Applications    2018, 54 (13): 1-10.   DOI: 10.3778/j.issn.1002-8331.1804-0271
    Abstract1087)      PDF(pc) (1072KB)(1932)       Save
    Road information plays an important role in modern society. It is of great scientific significance to study the road extraction method of remote sensing image. This paper reviews the development process of road extraction method, and divides the existing road extraction methods into three categories: based on pixel, object-oriented and deep learning according to the realization form. It is used as a clue to analyze and compare the scope of application of various methods with advantages and disadvantages. Design experiments, with a number of high-resolution satellite remote sensing images as the experimental object, verify the comparison of various types of typical road extraction method of the actual performance. The experimental results show that the method of road extraction based on deep learning is the best. Finally, based on the theory of popular remote sensing data and artificial intelligence, the development trend of road extraction method of remote sensing image is prospected.
    Related Articles | Metrics
    Survey of Data Fusion Based on Deep Learning
    ZHANG Hong, CHENG Chuanqi, XU Zhigang, LI Jianhua
    Computer Engineering and Applications    2020, 56 (24): 1-11.   DOI: 10.3778/j.issn.1002-8331.2007-0475
    Abstract724)      PDF(pc) (683KB)(1045)       Save

    As data fusion is the key to maximize the value of big data, while deep learning is a technical tool for mining deep characteristic information of data, data fusion based on deep learning can fully tap the potential value of big data, thus expanding the exploration and understanding of the world to a new depth and breadth. And this paper learns the advantages of deep learning in data fusion by reviewing the literature related to data fusion based on deep learning in recent years. The common data-fusion methods are classified, the advantages and disadvantages of which are pointed out. Analysis is conducted on data fusion method based on deep learning from three perspectives, namely the data fusion method extracted based on features of deep learning, data fusion method based on deep learning fusion and data fusion method based on the whole process of deep learning, and corresponding comparisons and summaries are conducted as well. This paper summarizes the whole document, discusses the difficulties in the application of deep learning in data fusion and the problems which require further research in the future.

    Related Articles | Metrics
    Computer Engineering and Applications    2020, 56 (24): 0-0.  
    Abstract197)      PDF(pc) (1126KB)(1322)       Save
    Related Articles | Metrics
    Design of Reward Function in Deep Reinforcement Learning for Trajectory Planning
    LI Yue, SHAO Zhenzhou, ZHAO Zhendong, SHI Zhiping, GUAN Yong
    Computer Engineering and Applications    2020, 56 (2): 226-232.   DOI: 10.3778/j.issn.1002-8331.1810-0021
    Abstract420)      PDF(pc) (1421KB)(525)       Save
    For the trajectory planning of robot manipulator in unknown environments, current deep reinforcement learning based?methods often suffer from the low learning efficiency and low robustness of planning strategy. To overcome the problems above, a novel azimuth reward function based trajectory planning method called A-DPPO is proposed. A novel azimuth reward function based on relative orientation and relative position is designed to reduce the invalid explorations and improve the learning efficiency. Moreover, it is the first time that Distributed Proximal Policy Optimization(DPPO) is applied to the trajectory planning for robot manipulator to improve the robustness of planning strategy. Experimental results show that the proposed A-DPPO method can increase the learning efficiency, compared to the state-of-the-art methods, and improve the robustness of planning strategy greatly.
    Related Articles | Metrics
    Survey of Intelligent Question Answering Research Based on Knowledge Graph
    WANG Zhiyue, YU Qing, WANG Nan, WANG Yaoguo
    Computer Engineering and Applications    2020, 56 (23): 1-11.   DOI: 10.3778/j.issn.1002-8331.2004-0370
    Abstract785)      PDF(pc) (774KB)(1043)       Save

    The answer selection model based on knowledge graph has become one of the hottest directions at present. This paper introduces the implementation of answer selection model based on knowledge graph from four aspects of template method, semantic parsing, deep learning and knowledge graph embedding, sums up their advantages, disadvantages and unsolved problem. Combined with the development of artificial intelligence technology, this paper introduces intelligent question-answer system based on deep learning. This research is helpful for more researchers to devote themselves to the intelligent question-answer system and develops different kinds of intelligent question-answer system to improve the social intelligent information service.

    Related Articles | Metrics
    Application of Deep Reinforcement Learning Algorithm on Intelligent Military Decision System
    KUANG Liqun, LI Siyuan, FENG Li, HAN Xie, XU Qingyu
    Computer Engineering and Applications    2021, 57 (20): 271-278.   DOI: 10.3778/j.issn.1002-8331.2104-0114
    Abstract146)      PDF(pc) (1223KB)(245)       Save

    Deep reinforcement learning algorithm can well achieve discrete decision-making behavior, but it is difficult to apply to the highly complex and continuous modern battlefield situations, and the algorithm is difficult to converge in multi-agent environment. To solve these problems, an improved Deep Deterministic Policy Gradient(DDPG) algorithm is proposed, which introduces the experience replay technology based on priority and single training mode to improve the convergence speed of the algorithm; at the same time, an exploration strategy of mixed double noise is designed in the algorithm to realize complex and continuous military decision-making and control behavior. The intelligent military decision simulation platform based on the improved DDPG algorithm is developed by unity3D. The simulation environment of Blue Army Infantry attacking Red Army military base is built to simulate multi-agent combat training. The experimental results show that the algorithm can drive multiple combat agents to complete tactical maneuvers and achieve tactical behaviors, such as bypassing obstacles to reach the dominant area for shooting. The algorithm has faster convergence speed and better stability. It can get higher round rewards, and achieves the purpose of improving the efficiency of intelligent military decision-making.

    Reference | Related Articles | Metrics
    Progress on Deep Reinforcement Learning in Path Planning
    ZHANG Rongxia, WU Changxu, SUN Tongchao, ZHAO Zengshun
    Computer Engineering and Applications    2021, 57 (19): 44-56.   DOI: 10.3778/j.issn.1002-8331.2104-0369
    Abstract537)      PDF(pc) (1134KB)(386)       Save

    The purpose of path planning is to allow the robot to avoid obstacles and quickly plan the shortest path during the movement. Having analyzed the advantages and disadvantages of the reinforcement learning based path planning algorithm, the paper derives a typical deep reinforcement learning, Deep Q-learning Network(DQN) algorithm that can perform excellent path planning in a complex dynamic environment. Firstly, the basic principles and limitations of the DQN algorithm are analyzed in depth, and the advantages and disadvantages of various DQN variant algorithms are compared from four aspects:the training algorithm, the neural network structure, the learning mechanism and AC(Actor-Critic) framework. The paper puts forward the current challenges and problems to be solved in the path planning method based on deep reinforcement learning. The future development directions are proposed, which can provide reference for the development of intelligent path planning and autonomous driving.

    Reference | Related Articles | Metrics
    Mixed Strategy Improved Sparrow Search Algorithm
    ZHANG Weikang, LIU Sheng, REN Chunhui
    Computer Engineering and Applications    2021, 57 (24): 74-82.   DOI: 10.3778/j.issn.1002-8331.2101-0161
    Abstract138)      PDF(pc) (1182KB)(223)       Save

    Aiming at the shortcomings of the sparrow search algorithm in the iterations of population diversity reduction, easy to fall into local optimality and slow convergence speed, a Mixed Strategy improved Sparrow Search Algorithm(MSSSA) is proposed. Circle map is used to initialize the individual positions of sparrows to increase the diversity of the initial population. Combining the butterfly optimization algorithm the location update method of the discoverer is improved to enhance global exploration ability of the algorithm. The dimensional-by-dimensional mutation method is used to perturb the individual position and improve the algorithm’s ability to jump out of the local optimum. In the simulation experiment, it compares with 4 basic algorithms and 5 improved algorithms based on 10 benchmark functions and performs Wilcoxon rank sum test. The results show that the proposed algorithm has better convergence and solution accuracy, global optimization ability has been greatly improved.

    Reference | Related Articles | Metrics
    MTICA-AEO-SVR Model for Stock Price Forecasting
    DENG Jiali, ZHAO Fengqun, WANG Xiaoxia
    Computer Engineering and Applications    2022, 58 (8): 257-263.   DOI: 10.3778/j.issn.1002-8331.2108-0433
    Abstract63)      PDF(pc) (2491KB)(196)       Save
    In order to improve the stability and separation efficiency of traditional Fast ICA algorithm, a new nonlinear function based on Tukey M estimation is constructed in this paper, and then a MTICA algorithm is obtained. Furthermore, a novel MTICA-AEO-SVR model for stock price forecasting is established combining MTICA and SVR algorithms. Firstly, the original stock data is decomposed into independent components by MTICA algorithm for sorting and denoising, and then different SVR models are selected to predict the independent components and the stock price respectively. Artificial ecosystem optimization is introduced into the SVR algorithm to select parameters, as to improve the model prediction accuracy. The empirical results of the Shanghai B-share index show that MTICA-AEO-SVR model is more accurate and efficient than ICA-AEO-SVR model and ICA-SVR model in stock price prediction.
    Reference | Related Articles | Metrics
    Application of Improved SSD Algorithm in Parts Detection
    SHEN Xinfeng, JIANG Ping, ZHOU Genrong
    Computer Engineering and Applications    2021, 57 (7): 257-262.   DOI: 10.3778/j.issn.1002-8331.2001-0097
    Abstract136)      PDF(pc) (1601KB)(345)       Save

    Aiming at the problems of high real-time and accuracy requirements of parts type detection during production, and the small volume of some parts is difficult to detect, a new method of parts detection based on the improved SSD target detection algorithm is proposed. The lightweight network MobileNetV3-Large is used to replace network VGG-16. Resizing image from 300×300 px to 224×224 px, the feature pyramid network is adopted to improve the detection effect of small parts. Taking the detection of pneumatic motor parts to test, during training, data enhancement is applied to improve the robustness of the model. The experimental results show that the improved SSD algorithm improves the speed of real-time detection of components and ensures detection accuracy.

    Related Articles | Metrics
    Review of Machine Learning for Predictive Maintenance
    LI Jieqi, HU Liangbing
    Computer Engineering and Applications    2020, 56 (21): 11-19.   DOI: 10.3778/j.issn.1002-8331.2006-0016
    Abstract338)      PDF(pc) (660KB)(883)       Save

    Machine learning algorithms can process high-dimensional and multi-variable data, and extract hidden relationships in the data in complex and dynamic environments, and have good application prospects in predictive maintenance technology. However, the performance of predictive maintenance system depends on the choice of machine learning algorithms. This paper reviews the current machine learning algorithms used in predictive maintenance system, compares the advantages and disadvantages of several machine learning algorithms characteristic in detail. The application of the machine learning in predictive maintenance is prospected in the future.

    Related Articles | Metrics
    Path Planning Using Improved RRT Algorithm for Indoor Mobile Robot
    LIU Ziyan,ZHANG Jie
    Computer Engineering and Applications    2020, 56 (9): 190-197.   DOI: 10.3778/j.issn.1002-8331.1903-0105
    Abstract262)      PDF(pc) (1149KB)(577)       Save

    Aiming at the defects of low sampling efficiency and high deviation from optimal solutions of basic RRT algorithm due to randomly selecting extended nodes, an improved RRT algorithm with goal-biased is proposed. After the extended nodes being selected by using the target bias strategy and the odor diffusion, random trees grow to target points. A path smoothing method based on B-spline curve is proposed, which has higher searching efficiency and path quality. The simulation results demonstrate that the path generated by the proposed algorithm is around 22.1% shorter than that of basic RRT algorithm and the path is smoother as well. Furthermore, the proposed algorithm has stronger ability of avoiding obstacles. Finally, the improved RRT algorithm is applied it to Turtlebot2 in real environment. The experimental results illustrate that the improved RRT algorithm achieves higher reliability and practicability.

    Related Articles | Metrics
    Multi-Scale Transformer Lidar Point Cloud 3D Object Detection
    SUN Liujie, ZHAO Jin, WANG Wenju, ZHANG Yusen
    Computer Engineering and Applications    2022, 58 (8): 136-146.   DOI: 10.3778/j.issn.1002-8331.2109-0489
    Abstract74)      PDF(pc) (1383KB)(188)       Save
    Point cloud 3D object detection has low detection accuracy for small objects such as pedestrians and bicycles, which is easy to miss detection and false detection. A 3D object detection method MSPT-RCNN(multi-scale point transformer-RCNN) based on multi-scale point cloud transformer is proposed to improve the detection accuracy of point cloud 3D objects. The method consists of two stages, the first stage(RPN) and the second stage(RCNN). In RPN stage, point cloud features are extracted through multi-scale transformer network, which includes multi-scale neighborhood embedding module and jump connection offset attention module to obtain multi-scale neighborhood geometric information and different levels of global semantic information, and generate high-quality initial 3D bounding box. In the RCNN stage, the multi-scale neighborhood geometric information of point cloud in the bounding box is introduced to optimize the position, size, orientation and confidence of the bounding box. The experimental results show that this method(MSPT-RCNN) has high detection accuracy, especially for distant and small objects. MSPT-RCNN can effectively improve the accuracy of 3D object detection by effectively learning the multi-scale geometric information in point cloud data and extracting different levels of effective semantic information.
    Reference | Related Articles | Metrics
    Survey of Small Object Detection Algorithms Based on Deep Learning
    LIU Yang, ZHAN Yinwei
    Computer Engineering and Applications    2021, 57 (2): 37-48.   DOI: 10.3778/j.issn.1002-8331.2009-0047
    Abstract915)      PDF(pc) (959KB)(917)       Save

    With the development of artificial intelligence technology, deep learning technology has been widely used in face recognition, pedestrian detection, unmanned driving and other fields. As one of the most basic and challenging problems in machine vision, object detection has attracted extensive attention in recent years. Aiming at the problem of object detection, especially small object detection, this paper summarizes the common data sets and performance evaluation metrics, and compares the characteristics, advantages and difficulties of various common data sets. At the same time, this paper systematically summarizes the common object detection methods and the challenges faced by small object detection. In addition, combing the latest work based on deep learning, this paper introduces the multi-scale and super-resolution small object detection methods in the highlight and presents the lightweight strategy and the performance of some lightweight models based on the object detection. Finally, this paper summarizes the characteristics, advantages and limitations of various methods, and looks at the future development direction of small object detection method based on deep learning.

    Related Articles | Metrics
    Channel and Spatial Attention Image Super Resolution Network
    LIU Jing, SONG Haichuan, HUANG Jianshe, MA Lizhuang
    Computer Engineering and Applications    2021, 57 (2): 209-216.   DOI: 10.3778/j.issn.1002-8331.1911-0296
    Abstract167)      PDF(pc) (1618KB)(400)       Save

    Single image super-resolution plays an important role in the field of computer vision. This technology aims to reconstruct high-resolution images from low-resolution images. In recent years, deep neural networks make performance in SISR task significantly improved. However, recently works based on convolutional neural network equally treat high-frequency and low-frequency features, which makes the reconstruction of high-frequency details poor, the output too smooth and the texture information lack. On the other hand, very deep convolutional network is not easy to converge, and as the depth of the neural network grows, the long-term information from the former layer can easily be weakened or lost in the latter layer, which makes the benefit not proportional to the depth of the network and the computational complexity. To solve these above problems, it proposes a spatial attention module and a channel attention module as the basic block of convolutional neural network for SISR. Firstly, in the same channel, the information of different locations is given different weights by the spatial attention module. Secondly, the weights between different channels are determined by the channel attention module, which makes the high-frequency information gain a higher position in the reconstruction task. The reconstruction performance is improved. It further proposes a short-term and long-term feature modulation module to transform the layer depth of the network into the block depth, which greatly reduces the depth of the network, in order to solve the problem of long-term information loss in the front layer. Compared with other methods based on deep convolution neural network, the Peak Signal-to-Noise Ratio(PSNR) on several benchmark datasets are better, which proves the effectiveness of the proposed method.

    Related Articles | Metrics
    Real-Time Heart Rate Extraction Method of Anti-Motion Interference
    YANG Gang, XUE Ting, GAO Wei
    Computer Engineering and Applications    2019, 55 (5): 244-250.   DOI: 10.3778/j.issn.1002-8331.1711-0166
    Abstract311)      PDF(pc) (758KB)(1667)       Save
    When extracting motion heart rate on the base of PPG, measurement result of conventional heart-rate extraction algorithm faces large error and bad real-time performance problems due to motion noise interference, a real-time heart rate extraction method of anti-motion interference is then presented. By using real-time wavelet de-noising, classifying and then training motion by combining ACC, heart rate gain of each motion status is calculated and real-time heart rate is compensated in the method. It is indicated by experimental result that, when compared with real-time heart rate calculated by using EGG signals collected at the same time, the absolute error rate is just around 1.2%. When compared with conventional heart rate extraction algorithm, the method is characterized with strong anti-interference and accurate real-time performance.
    Related Articles | Metrics
    Survey of Image Compression Algorithm Based on Deep Learning
    YU Heng, MEI Hongyan, XU Xiaoming, JIA Huiping
    Computer Engineering and Applications    2020, 56 (15): 15-23.   DOI: 10.3778/j.issn.1002-8331.2003-0294
    Abstract481)      PDF(pc) (923KB)(939)       Save

    With the continuous development of deep learning and the explosive growth of image data, how to use deep learning to obtain higher compression ratio and higher quality images has gradually become one of the hot research issues. Through the analysis of the related literatures in recent years, the image compression method based on the deep learning is summarized and analyzed according to the Convolutional Neural Network(CNN), Recurrent Neural Network(RNN), Generative Adversarial Network(GAN). This paper enumerates the typical examples, and the image compression algorithm based on depth study of the training data set, commonly used evaluation indexes are introduced, according to the deep learning advantages in the field of image compression for its future development trend are summarized and discussed.

    Related Articles | Metrics
    Attention Mechanism-Based CNN-LSTM Model and Its Application
    LI Mei1,2, NING Dejun1, GUO Jiacheng1,2
    Computer Engineering and Applications    2019, 55 (13): 20-27.   DOI: 10.3778/j.issn.1002-8331.1901-0246
    Abstract862)      PDF(pc) (914KB)(2219)       Save
    Time series have temporal property, and the characteristics of its short sequences are different in importance. Aiming at the characteristics of time series, a neural network prediction model based on Convolution Neural Network(CNN) and Long Short-Term Memory(LSTM) is proposed, which combines coarse and fine grain features to achieve accurate time series prediction. The model consists of two parts. CNN based on attention mechanism adds attention branch to standard CNN network to extract important fine-grained features. The back end is LSTM, which extracts the coarse-grained features of the hidden time series from fine-grained features. Experiments on real cogeneration heat load dataset demonstrate that the model is better than the autoregressive integrated moving average, support vector regression, CNN and LSTM models. Compared with the pre-determined method currently used by enterprises, the Mean Absolute Scaled Error(MASE) and Root Mean Square Error(RMSE) have been increased by 89.64% and 61.73% respectively.
    Related Articles | Metrics
    Research on Image Style Transfer Technology Based on Semantic Segmentation
    LI Meili, YANG Chuanying, SHI Bao
    Computer Engineering and Applications    2020, 56 (24): 207-213.   DOI: 10.3778/j.issn.1002-8331.1910-0238
    Abstract263)      PDF(pc) (1172KB)(390)       Save

    With the collision and fusion of national costume culture, this paper studies the image style transfer technology, expounds the current research status of style transfer, integrates Mongolian costume style with Han style, and inherits and promotes the national culture. For large difference of Mongolian costume elements variety, color, decorative pattern characteristics and cause of style such as irregularity extraction is difficult problem, it uses the algorithm of [K]-means and closed natural cutout combination method for image segmentation, extracts the image of style and content based on neural network, uses image reconstruction technology to synthesize results, implements the image style transfer of Mongolian and Han clothing. According to the serious output image artifact, it adopts the migration algorithm, an improved image style will constrain the transform of the input image to the output image in the local affine transformation of color space, the constraints are represented as a differentiable parameter completely, it effectively restrains image distortion, at the same time in real style photos do not match the space problems in the process of migration, it treatments smoothly to ensure the space style is consistent, this method greatly accelerates the speed.

    Related Articles | Metrics
    Review of Research on Generative Adversarial Networks and Its Application
    WEI Fuqiang, Gulanbaier Tuerhong, Mairidan Wushouer
    Computer Engineering and Applications    2021, 57 (19): 18-31.   DOI: 10.3778/j.issn.1002-8331.2104-0248
    Abstract386)      PDF(pc) (1078KB)(994)       Save

    The theoretical research and applications of generative adversarial networks have been continuously successful and have become one of the current hot spots of research in the field of deep learning. This paper provides a systematic review of the theory of generative adversarial networks and their applications in terms of types of models, evaluation criteria and theoretical research progress; analyzing the strengths and weaknesses of generative models with explicit and implicit density-based, respectively; summarizing the evaluation criteria of generative adversarial networks, interpreting the relationship between the criteria, and introduces the research progress of the generative adversarial network in image generation from the application level, that is, through the image conversion, image generation, image restoration, video generation, text generation and image super-resolution applications; analyzing the theoretical research progress of generative adversarial networks from the perspectives of interpretability, controllability, stability and model evaluation methods. Finally, the paper discusses the challenges of studying generative adversarial networks and looks forward to the possible future directions of development.

    Reference | Related Articles | Metrics
    Survey on Vision-Based 3D Object Detection Methods
    LI Yujie, LI Xuanpeng, ZHANG Weigong
    Computer Engineering and Applications    2020, 56 (1): 11-24.   DOI: 10.3778/j.issn.1002-8331.1909-0024
    Abstract460)      PDF(pc) (825KB)(1390)       Save
    Vision-based object detection is an important component of environment perception systems. It has been a research hotspot in computer vision, robotics and other related fields. The 3D object detection is based on the 2D object detection, which involves the estimation of the object scale, localization and pose estimation in the camera coordinate. Compared to 2D object detection, there are still a big gap for 3D object detection in terms of accuracy and real-time performance. This paper systematically surveys the state-of-the-art vision-based 3D object detection methods based on monocular vision, stereo vision and RGB-D, and classifies them according to indoor and outdoor scenes. In addition, the paper compares and analyzes these methods on KITTI, SUN RGB-D and other datasets, and discusses on the future research direction.
    Related Articles | Metrics
    Survey of Network Traffic Forecast Based on Deep Learning
    KANG Mengxuan, SONG Junping, FAN Pengfei, GAO Bowen, ZHOU Xu, LI Zhuo
    Computer Engineering and Applications    2021, 57 (10): 1-9.   DOI: 10.3778/j.issn.1002-8331.2101-0402
    Abstract900)      PDF(pc) (711KB)(820)       Save

    Precisely predicting the trend of network traffic changes can help operators accurately predict network usage, correctly allocate and efficiently use network resources to meet the growing and diverse user needs. Taking the progress of deep learning algorithms in the field of network traffic prediction as a clue, this paper firstly elaborates the evaluation indicators of network traffic prediction and the current public network traffic data sets. Secondly, this paper specifically analyzes four deep learning methods commonly used in network traffic prediction:deep belief networks, convolutional neural network, recurrent neural network, and long short term memory network, and focuses on the integrated neural network models used in recent years for different problems. The characteristics and application scenarios of each model are summarized. Finally, the future development of network traffic forecast is prospected.

    Related Articles | Metrics
    Improved Leukocyte Detection Algorithm of YOLOv5
    WANG Jing, SUN Ziyun, GUO Ping, ZHANG Longmei
    Computer Engineering and Applications    2022, 58 (4): 134-142.   DOI: 10.3778/j.issn.1002-8331.2107-0332
    Abstract131)      PDF(pc) (1380KB)(175)       Save
    Aiming at the problems of low accuracy and poor effect caused by small white blood cell data samples, small difference between classes and small target size, this paper proposes a white blood cell detection algorithm YOLOv5-CHE based on improved YOLOv5. Firstly, coordinate attention mechanism is added to the convolutional layer of the backbone feature extraction network to improve the feature extraction capability of the algorithm. Secondly, the purpose of using four-scale feature detection and reacquiring anchor point frame is to increase the detection scale of shallow layer and improve the recognition accuracy of small targets. Finally, the purpose of changing the bounding box regression loss function is to improve the accuracy of check box detection. Experimental results show that the mean average precision(mAP), precision and recall of YOLOv5-CHE are improved by 3.8 percentage points, 1.8 percentage points and 1.5 percentage points in comparison with the benchmark YOLOv5 algorithm, respectively, which shows that the proposed algorithm is effective for leukocyte detection.
    Reference | Related Articles | Metrics
    License Plate Location Detection Algorithm Based on Improved YOLOv3 in Complex Scenes
    MA Qiaomei, WANG Mingjun, LIANG Haoran
    Computer Engineering and Applications    2021, 57 (7): 198-208.   DOI: 10.3778/j.issn.1002-8331.2008-0137
    Abstract185)      PDF(pc) (1709KB)(549)       Save

    Aiming at the problem of the difficulty of license plate positioning, slow detection speed and low detection accuracy in complex scenes such as lighting, multi-vehicle and low resolution, an improved method based on YOLOv3 is proposed. Firstly, the label information of the example is clustered by K-means++ method to obtain a new anchor size. And then, the improved thin feature extraction network(DarkNet41) is used to improve the detection efficiency of the model and reduce computational consumption. Moreover, multi-scale feature fusion is improved from 3-scale prediction to 4-scale prediction and improved Inception-SE structure is added to the detection network to improve the accuracy of detection. Finally, CIoU is selected as a loss function. The data is enhanced with the Multi-Scale Retinex(MSR) algorithm. Experimental analysis shows that the improved algorithm’s mAP reaches 98.84% and the detection speed reaches 36.4 frame/s, which has better accuracy and real-time performance compared with the YOLOv3 model and other algorithms.

    Related Articles | Metrics
    Application of Improved XGBoost Model in Stock Forecasting
    WANG Yan, GUO Yuankai
    Computer Engineering and Applications    2019, 55 (20): 202-207.   DOI: 10.3778/j.issn.1002-8331.1904-0007
    Abstract327)      PDF(pc) (1003KB)(1089)       Save
    With the continuous advancement of the times, people’s living standards have been increasing. In addition to solving the problem of food and clothing, there is surplus money available for investment. More and more people are turning their attention to stock market investment, which provides financial conditions for the development of the stock market. However, in the complicated stock market, how to find the optimal stock has become an urgent problem to be solved. This is not only a unilateral confusion for investors, but also a focus of scholars in the field of stock forecasting. In this paper, the grid prediction algorithm is used to optimize the XGBoost model to construct the financial forecasting model of GS-XGBoost, and the model is applied to short-term stock forecasting. The daily closing prices of China Ping An, China State Construction Engineering Corporation, CRRC Corporation Limited, IFLYTEK and SANY HEAVY INDUSTRY from April 2005 to December 28, 2018 are used as experimental data. Through experimental comparison, compared with the original XGBoost model, GBDT model and SVM model, the GS-XGBoost model shows good prediction results on the three evaluation indexes of MSE, RMSE and MAE. It is verified that the GS-XGBoost financial forecasting model has better fitting performance in short-term stock forecasting.
    Related Articles | Metrics
    Survey of Knowledge Reasoning Based on Neural Network
    ZHANG Zhongwei1,2, CAO Lei1, CHEN Xiliang1, KOU Dalei1,3, SONG Tianting2
    Computer Engineering and Applications    2019, 55 (12): 8-19.   DOI: 10.3778/j.issn.1002-8331.1901-0358
    Abstract554)      PDF(pc) (777KB)(2657)       Save
    Knowledge reasoning is an important means of knowledge graph completion and has always been one of the research hotspots in the field of knowledge graph. With the development of neural network, its applications in knowledge reasoning have been paid more and more attention in recent years. The knowledge reasoning methods based on neural network have not only stronger reasoning and generalization abilities, but also higher utilization rates of entities, attributes, relations and text information in the knowledge base. These methods are more effective in reasoning. The relevant concepts of knowledge graph and knowledge graph completion are introduced, the concepts and basic principles of knowledge reasoning are indicated, and then the latest research progresses of the technology of knowledge reasoning based on neural network are reviewed. The existing problems and development directions of knowledge reasoning in the aspect of theory, algorithm and application are summarized.
    Related Articles | Metrics
    Field Weed Identification Method Based on Deep Connection Attention Mechanism
    SHU Yali, ZHANG Guowei, WANG Bo, XU Xiaokang
    Computer Engineering and Applications    2022, 58 (6): 271-277.   DOI: 10.3778/j.issn.1002-8331.2108-0077
    Abstract85)      PDF(pc) (1037KB)(167)       Save
    In order to achieve fast and accurate recognition of field weed images, a field weed recognition model based on deep connected attention mechanism residual network(DCECA-Resnet50-a) is proposed. Using the residual network as a benchmark, this paper improves the position of residual block downsampling, introduces the attention mechanism and connected attention mechanism modules to better extract the feature information in the images, combines the migration learning strategy to alleviate the overfitting phenomenon caused by small sample data sets, improves the generalization of the model and greatly reduces the training time of the model. The experimental results show that the improved model has the best overall performance and high recognition accuracy, with 96.31% accuracy for weeds and fewer model parameters, and achieves the accurate differentiation of four types of common weeds in pea fields, namely, silverleaf daisy, chaparral, matang and pigweed, which provides a corresponding reference for small sample data in the field of agricultural recognition.
    Reference | Related Articles | Metrics
    Ranking Algorithms of Vital Nodes Based on Spring Model
    MENG Yuyu, WANG Xiao, YAN Guanghui, LUO Hao, YANG Bo, ZHANG Lei, WANG Qiong
    Computer Engineering and Applications    2022, 58 (7): 77-86.   DOI: 10.3778/j.issn.1002-8331.2103-0204
    Abstract63)      PDF(pc) (1393KB)(167)       Save
    Node ranking of vital nodes is an important problem in complex networks. When using the robustness and vulnerability of the network to evaluate the node ranking algorithms gravity model (GM) and local gravity model (LGM) based on the gravity model, once the nodes with large degrees have been removed from the network, the removal of neighbors with large gravitational values usually cannot largely affect the structure and function of the network, which shows that the algorithms still have some improvement in the ranking accuracy of vital nodes. Because of that, inspired by the spring model, further considering neighbors’ information and path information in the network, combined with the network diameter, spring model(SM) and local spring model(LSM), the node ranking algorithm and its local algorithm, are proposed. The results show that the SM algorithm and the LSM algorithm have higher accuracy for the ranking of vital nodes than other classical algorithms in synthetic networks and real networks. Especially, the SIR epidemic experiments on the Power network are conducted to furtherly verify the higher rationality and effectiveness of the SM algorithm than other algorithms.
    Reference | Related Articles | Metrics
    Summary of Feature Selection Methods
    LI Zhiqin, DU Jianqiang, NIE Bin, XIONG Wangping, HUANG Canyi, LI Huan
    Computer Engineering and Applications    2019, 55 (24): 10-19.   DOI: 10.3778/j.issn.1002-8331.1909-0066
    Abstract452)      PDF(pc) (964KB)(1455)       Save
    As a data preprocessing process, feature selection plays an important role in data mining, pattern recognition and machine learning. Through feature selection, the complexity of the problem can be reduced, and the prediction accuracy, robustness and interpretability of the learning algorithm can be improved. This paper introduces the framework of feature selection methods, and focuses on the two processes of generating feature subsets and evaluation criteria. The feature selection algorithms are classified according to different combinations of feature selection and learning algorithms, and the advantages and disadvantages of various methods are analyzed. The existing problems of existing feature selection algorithms are discussed, and some research difficulties and research directions are proposed.
    Related Articles | Metrics