Computer Engineering and Applications

Select

Research Progress on Multi-Agent Deep Reinforcement Learning and Scalability

LIU Yanfei, LI Chao, WANG Zhong, WANG Jieling

Computer Engineering and Applications 2025, 61 (4): 1-24. DOI: 10.3778/j.issn.1002-8331.2407-0034

Abstract （193）

PDF（pc）（2161KB）（223）

Save

Multi-agent deep reinforcement learning has shown great potential in solving agent collaboration, competition, and communication problems in recent years. However, as its application expands across more domains, scalability has become a focal concern, which is an important problem from theoretical research to large-scale engineering applications. This paper reviews the reinforcement learning theory and typical algorithms of deep reinforcement learning, introduces three learning paradigms of multi-agent deep reinforcement learning and their representative algorithms, and briefly summarizes the current mainstream open-source experimental platforms. Then, this paper delves into the research progress on the scalability of the number and scenarios in multi-agent deep reinforcement learning, analyzes the main problems faced by each method and providing existing solutions. Finally, the application prospect and development trend of multi-agent deep reinforcement learning are prospected, providing references and inspiration to further advance research in this field.

Reference | Related Articles | Metrics

Select

Review of Lung CT Image Lesion Region Segmentation Based on Deep Learning

LI Xiaotong, MA Sufen, SHENG Hui, WEI Guohui, LI Xintong

Computer Engineering and Applications 2025, 61 (4): 25-42. DOI: 10.3778/j.issn.1002-8331.2403-0315

Abstract （170）

PDF（pc）（4394KB）（191）

Save

Lung cancer poses a serious threat to people’s lives and health. The morphology of lesion areas in lung CT images is complex and diverse, and achieving high-precision segmentation of lesion areas in lung CT images has become a highly challenging key issue in the field of computer-aided diagnosis. The segmentation of lung lesion regions based on deep learning not only helps doctors diagnose early lung cancer quickly and accurately, but also has important clinical value for the treatment of lung cancer. In order to conduct in-depth research on lung lesion segmentation techniques, common datasets and evaluation indicators are introduced. The deep learning lung lesion regions segmentation models are reviewed in three aspects：segmentation model based on convolutional neural network, segmentation model based on U-Net model, and segmentation model based on generative adversarial network. The innovative points of domestic and foreign research over the past 5 years are summarized through specific experiments. The segmentation performance of various models is compared and analyzed. The advantages and disadvantages of various models are summarized, and the development direction in this field is discussed.

Reference | Related Articles | Metrics

Select

Improved Lightweight Multi-Directional License Plate Detection Algorithm of YOLOX

LEI Jingsheng, ZHANG Zhihao, QIAN Xiaohong, WANG Weiran, YANG Shengying

Computer Engineering and Applications 2025, 61 (4): 230-240. DOI: 10.3778/j.issn.1002-8331.2309-0283

Abstract （97）

PDF（pc）（1507KB）（117）

Save

In response to the problems of the existing license plate detection algorithms in complex environments, such as poor performance in detecting multi-directional license plates, low real-time capabilities, and excessive model parameters and computational complexity, a lightweight multi-directional license plate detection algorithm based on YOLOX is proposed. By adjusting the number of residual components and using a combination of large convolution kernels and depthwise separable convolutions, the parameter count of the backbone network is reduced. A channel attention mechanism is introduced to effectively extract channel interaction information and reduce noise interference. The feature fusion network is lightweighted by using depthwise separable convolutions and adjusting the expansion ratio. A rotation decoupling head is designed, and an angle prediction branch is added to enable more accurate prediction of the rotation bounding boxes of multi-directional license plates. The rotation IoU loss is used instead of the horizontal IoU loss to improve detection accuracy. Experimental results on the CCPD dataset show that the improved algorithm has the parameter count and computational complexity of 2.38 million and 12.97 GFLOPs, respectively, which are reduced by 45% and 33% compared to YOLOX-tiny. The detection accuracy AP70 is 94.9%, and the detection frame rate is 76.6 FPS. The improved license plate detection model can detect multi-directional license plates in real-time while maintaining high accuracy.

Reference | Related Articles | Metrics

Select

Crack Detection Method Based on Attention Decoder and Continuous Supervision

XIE Yonghua, ZHUO Annan

Computer Engineering and Applications 2025, 61 (4): 122-129. DOI: 10.3778/j.issn.1002-8331.2309-0292

Abstract （74）

PDF（pc）（2197KB）（88）

Save

Crack detection is an important measure to prevent building collapse accidents. Aiming at the weak crack feature and interference from noise in crack images, encoder-decoder nets in semantic segmentation are introduced to improve the detection robustness. An attention decoder module which enhances the attention of the crack feature in decoder through attention mechanism is designed to address the shortcomings of the simple decoder in traditional encoder-decoder nets and the low efficiency of connecting the semantic information in encoding features. For the problem of cracks fracture in predictions, the continuous supervision algorithm is improved, by adding adjacency connected predictions and combining the corresponding loss function for supervision, the continuity of crack features in prediction results is improved. Combining these two methods, an automatic encoder-decoder crack detection model is proposed. The experimental results on two datasets verify the superiority of the proposed model, the best detection accuracy is achieved among the commonly used models selected.

Reference | Related Articles | Metrics

Select

Deep Reinforcement Learning Navigation Algorithm for Coexisting-Cooperative-Cognitive Robots in Dynamic Environment

GU Jinhao, KUANG Liqun, HAN Huiyan, CAO Yaming, JIAO Shichao

Computer Engineering and Applications 2025, 61 (4): 90-98. DOI: 10.3778/j.issn.1002-8331.2405-0088

Abstract （82）

PDF（pc）（1106KB）（86）

Save

In the past few decades, navigation algorithms for mobile service robots have been extensively studied, but intelligent agents still lack the complexity and cooperation exhibited by humans in crowded environments. With the continuous expansion of human-machine integration applications, collaboration between robots and humans in shared workspaces will become increasingly important. Therefore, the next generation of mobile service robots needs to meet social requirements in order to be accepted by humans. In order to enhance the autonomous navigation ability of multi-agent systems in dynamic scenarios, a deep reinforcement learning obstacle avoidance algorithm for coexisting-cooperative-cognitive robots in dynamic environments is proposed to address the issues of low social adaptability and finding the optimal value function in multi-agent navigation. A motion model that is more closely related to human behavior is established and added to the deep reinforcement learning framework to improve the cooperation of coexisting-cooperative-cognitive robots. In order to enhance the perceived safety of pedestrians based on their physical safety, a reward function is redefined. Nonlinear deep neural networks are used instead of traditional value functions to solve the problem of finding the optimal value function. Simulation experiments show that compared to the latest deep reinforcement learning navigation methods, the proposed method achieves a 100% navigation success rate without increasing navigation time and without any collisions. The results indicate that this method maximizes the satisfaction of human social principles for the fusion robot, while effectively moving towards the goal and improving the perceived safety of pedestrians.

Reference | Related Articles | Metrics

Select

Survey on Applications of Deep Reinforcement Learning in Edge Video Transmission Optimization

LI Yan, WAN Zheng

Computer Engineering and Applications 2025, 61 (4): 43-58. DOI: 10.3778/j.issn.1002-8331.2402-0156

Abstract （65）

PDF（pc）（1201KB）（85）

Save

In the era of industrial video, with the rapid development of edge computing and artificial intelligence technology, the edge intelligence has been born, and the research of video transmission optimization based on edge computing network has ushered in new opportunities. Based on the summary of the content of edge video transmission optimization, this paper expounds the research and progress of deep reinforcement learning applied to edge video transmission optimization. The concept of edge intelligent video transmission optimization is proposed, and the method classification of edge intelligent video transmission optimization is presented. It consists of edge intelligent video transmission optimization for network QoS and edge intelligent video transmission optimization for user QoE, and the detailed description is respectively given. The main problems of edge video transmission optimization at present are studied and analyzed, the future hot research directions of edge intelligent video transmission optimization are pointed out by finding laws, summarizing deficiencies and highlighting advantages.

Reference | Related Articles | Metrics

Select

Computer Engineering and Applications 2025, 61 (4): 0-0.

Abstract （74）

PDF（pc）（696KB）（81）

Save

Related Articles | Metrics

Select

LMUAV-YOLOv8: Lightweight Network for Object Detection in Low-Altitude UAV Vision

DONG Yibing, ZENG Hui, HOU Shaojie

Computer Engineering and Applications 2025, 61 (3): 94-110. DOI: 10.3778/j.issn.1002-8331.2407-0127

Abstract （100）

PDF（pc）（4352KB）（112）

Save

To tackle the challenges of weak sensing capacity and high missed detection rates for small-scale objects using low-altitude UAV in complex traffic scenarios, the LMUAV-YOLOv8 algorithm is proposed. Its efficiency and advantage are verified through ablation and comparative experiments. The internal mechanisms is visualized by using the method of class activation mapping. In this dissertation, a lightweight feature fusion network (UAV_RepGFPN) is introduced firstly, proposing new feature fusion paths and a feature fusion module DBB_GELAN, which reduces the number of parameters and computation while improving the performance of the feature fusion network. Secondly, the feature extraction module (FTA_C2f) is constructed using partial convolution (PConv) and triplet attention mechanism (Triplet Attention), and the ADown down-sampling module is introduced. By rearranging the dimensions of the input feature maps and making fine-grained adjustments, the ability of the deep network to capture spatial features is enhanced, further reducing the number of parameters and computation. Then, concerning large amount of information loss during in layer-by-layer feature extraction and spatial transformation, a new context-guided programmable gradient information (UAV_PGI) strategy is proposed. By designing a context-guided reversible architecture and three additional auxiliary detection heads, UAV_PGI significantly enhance detection capabilities for aerial objects. In order to verify the validity and generalization ability of the model, comparative experiments are carried out on the VisDrone 2019 test set, and the results show that: compared with YOLOv8s, LMUAV-YOLOv8s on the VisDrone 2019 test set improves precision, recall, mAP@0.5, and mAP@0.5:0.95 by 4.2, 3.9, 5.1, and 3.0?percentage points, separately, with the computational cost increased by only 0.4?GFLOPs and the parameter count reduced by 63.9%, meaning a good balance between performance and cost. The inference experimental results based on NVIDIA Jetson Xavier NX embedded platform show that compared with the baseline model, the proposed algorithm can obtain higher detection accuracy under the condition of meeting the requirements of real-time detection, rendering it more suitable for real-time target detection scenarios in drones. Finally, the decision making process is visualized by using the method of class activation mapping, which provides a intuitive way to understand the internal mechanisms of the networ. And the results show that the proposed model has superior small-scale feature extraction and high-resolution processing capabilities.

Reference | Related Articles | Metrics

Select

PMM-YOLO：Traffic Sign Detection Algorithm with Multi-Scale Feature Fusion

ZHAO Lei, LI Dong

Computer Engineering and Applications 2025, 61 (4): 262-271. DOI: 10.3778/j.issn.1002-8331.2405-0103

Abstract （89）

PDF（pc）（1613KB）（75）

Save

Traffic signs play a crucial role in the field of autonomous driving. However, they often present challenges such as small size, susceptibility to occlusion, and missed detections and false alarms in complex environments. This paper proposes a PMM-YOLO traffic sign detection algorithm based on improvements to YOLOv5. To effectively extract multi-scale information and enhance the model’s feature representation capability, an adaptive parallel atrous convolution (APA) module combining attention mechanism is introduced. Utilizing parallel atrous convolutions with different dilation rates enables effective extraction of features at various scales, while a gate mechanism highlights the representation of key targets, improving detection accuracy. A multi-branch adaptive sampling (MBAS) approach is designed to provide multiple feature extraction pathways for the network, enriching feature expression diversity. The important features are reinforced by the weight at different positions, and redundant features are suppressed. A multi-scale feature fusion (MSFF) module is devised to concatenate feature maps of different sizes, leveraging multi-scale information to fuse feature maps of multiple scales comprehensively, thus obtaining more comprehensive target features and enhancing detection performance. An output reorganization (ORO) module is constructed to enhance the detection of small targets by adding a small target detection layer and removing the large target detection layer, thereby reducing model complexity accordingly. Experimental results demonstrate that the PMM-YOLO algorithm achieves an mAP@0.5 of 86.4% on the TT100K dataset, representing a 5.9 percentage points improvement over the original YOLOv5. Additionally, the FPS is increased by 4.4% compared to the baseline, enabling rapid and accurate detection of traffic signs.

Reference | Related Articles | Metrics

Select

Review of Collaborative Inference Methods for Edge Intelligence

ZHAO Chanchan, LYU Fei, SHI Bao, YU Xiaomin, YANG Xingchen, YUE Xiaocan

Computer Engineering and Applications 2025, 61 (3): 1-20. DOI: 10.3778/j.issn.1002-8331.2406-0040

Abstract （126）

PDF（pc）（7788KB）（163）

Save

With the development of edge intelligence, collaborative inference technology has made significant progress in enhancing the efficiency and performance of intelligent applications through collaboration among cloud, edge, and terminal devices. The performance metrics, application scenarios, and challenges of edge intelligence are outlined, introducing four inference paradigms under collaborative inference technology through the rating architecture of edge intelligence: end-to-end collaboration, edge-to-end collaboration, edge-to-edge collaboration, and cloud-edge-end collaboration. Based on the limitations and differences of application scenarios for collaborative inference technology, the advantages, limitations, principles, and optimization goals of collaborative inference technology in different inference paradigms are comprehensively analyzed and compared. The discussion delves into issues such as computational resource allocation, inference latency optimization, and throughput optimization solved by collaborative inference technology in different application scenarios. It also points out challenges in privacy security, communication service resource management, and collaborative training within edge intelligence. Future development trends and research directions are discussed, providing references and insights for research in this field.

Reference | Related Articles | Metrics

Select

Cross-Modal Transparent Object Segmentation Combining CNN-Transformer

PAN Weilan, ZHANG Rongfen, LIU Yuhong, ZHANG Jiyou, SUN Long

Computer Engineering and Applications 2025, 61 (4): 222-229. DOI: 10.3778/j.issn.1002-8331.2310-0064

Abstract （61）

PDF（pc）（997KB）（72）

Save

Transparent objects have visual characteristics such as high transparency, glossiness and special texture, which make the boundary between the object and the background often blurred, making it difficult for traditional image segmentation algorithms to accurately recognize and segment them, so this paper proposes a cross-modal semantic segmentation algorithm for transparent objects, CTNet, which combines CNN-Transformer. The algorithm adopts the encoding-decoding structure of CNN and Transformer hybrid network to predict the category and location of transparent objects across modalities, CNN is used to extract image features, and Transformer is used for multimodal fusion transformer (MFT). The enhanced boundary attention module (EBAM) is designed to improve the image edge segmentation ability. A multi-scale fusion decoding structure is proposed to reduce the blurred features. The mean absolute error (MAE) of CTNet is 3.3% in the RGB-T-Glass dataset, and the intersection over union (IOU) is 90.18% and 95.00% in the test sets with transparent objects and without transparent objects, respectively. On the GDD dataset, the MAE is 6.9% and the IOU is 87.6%. The results show that CTNet successfully realizes accurate segmentation of transparent objects using visible and thermal infrared images, and meets the requirements of accuracy and robustness when segmenting transparent objects in the target task.

Reference | Related Articles | Metrics

Select

Improved RT-DETR Algorithm for Aerial Small Object Detection

LIU Siyuan, GAO Kai, YONG Longquan

Computer Engineering and Applications 2025, 61 (4): 272-281. DOI: 10.3778/j.issn.1002-8331.2407-0399

Abstract （75）

PDF（pc）（1975KB）（70）

Save

Aiming to address the issue of missed and false detection of small objects in aerial photography images by existing object detection algorithms, an improved algorithm based on RT-DETR (real-time detection transformer) is proposed. Partial convolution (PConv) is introduced into the backbone network, and a PConvBlock structure is designed. Then, a BasicBlock-PConvBlock module composed of PConvBlocks replaces the original BasicBlock, effectively reducing the number of model parameters. The bidirectional feature pyramid network (BiFPN) structure is adopted to optimize the feature fusion module. The S2 feature is introduced to enhance the detection ability of small objects. The CARAFE upsampling operator is introduced to strengthen the fast fusion of multi-scale features. Experimental results show that the improved model has a 13.9% reduction in parameter number compared to the RT-DETR model, and the mAP0.5 and mAP0.5：0.95 indicators are improved by 2.4 and 1.9 percentage points, respectively on the VisDrone test set. On the TT100K and DOTA datasets, the improved model outperforms the RT-DETR algorithm. The improved model significantly enhances detection accuracy while maintaining a smaller parameter number and computational cost, meeting the real-time detection application requirements for drone aerial photography images.

Reference | Related Articles | Metrics

Select

Improved Lightweight and Efficient FMG-YOLOv8s Algorithm for Steel Surface Defect Detection

LIANG Liming, LONG Pengwei, LI Yulin

Computer Engineering and Applications 2025, 61 (3): 84-93. DOI: 10.3778/j.issn.1002-8331.2406-0358

Abstract （113）

PDF（pc）（4445KB）（122）

Save

In response to the current challenges of low efficiency and poor accuracy in steel surface defect detection, as well as the complexity, large parameter size, and subpar detection accuracy and real-time performance of existing defect detection models, this paper proposes a lightweight and efficient steel defect detection algorithm (FMG-YOLOv8s) based on the YOLOv8s model. This method utilizes the lightweight FasterNet network as the backbone network to reduce model complexity and better handle multi-scale feature information for improved detection performance. The feature interaction module (M-C2f) is restructured to effectively preserve spatial and channel features, suppress redundant information, and enhance detection accuracy and speed. The GS-Detect module is designed as the detection network of the overall model to reduce model complexity and enhance training and inference speed. Experimental validation on the Severstal steel defect dataset shows that compared to the YOLOv8s algorithm, the FMG-YOLOv8 algorithm achieves a 3.3 percentage points improvement in mAP, while reducing parameter size and computational complexity by 8.2×106 and 2.21×1010 respectively, and reaching a detection speed of 250 frames per second with a 6.9 percentage points improvement in recall rate. Experimental results demonstrate that this algorithm strikes a better balance in terms of detection accuracy, speed, and lightweight design, providing reliable references for high-precision, lightweight, and real-time performance on edge-terminal devices. Generalization validation on the NEU-DET defect dataset shows a 3.1 percentage points improvement in mAP and 185 frames per second improvement in detection speed compared to the original model, indicating good robustness of the algorithm.

Reference | Related Articles | Metrics

Select

RCNN Method of Transmission Tower Component Detection Based on Knowledge Graph and Small Object Improvement

ZHANG Kai, JIA Tao

Computer Engineering and Applications 2025, 61 (4): 299-309. DOI: 10.3778/j.issn.1002-8331.2309-0381

Abstract （50）

PDF（pc）（5883KB）（63）

Save

Electric power inspection is an important part of transmission line construction. Using drones to inspect transmission towers and using deep learning technology to assist technicians in making intelligent decisions, can reduce false detection rate and improve detection efficiency. Existing studies are mostly incapable of fully recognizing tower components from all perspectives and scales, or adapting to the complex scenes of transmission tower images. To solve these issues, an RCNN method of transmission tower component detection based on knowledge graph and small object improvement is proposed. Firstly, a spatial knowledge graph module is constructed based on the Reasoning-RCNN model to model the spatial relationships among the detected boxes in the image. Then, an ROI context feature fusion module is constructed to address the small object problem, and a small object detection strategy based on image partitioning is introduced. The image data of transmission tower are manually annotated and the proposed method is evaluated on this dataset. The experimental results show that the proposed method achieves full-scale detection of transmission tower components in complex scenes. The comparison results also demonstrate the superior performance of the proposed method over baseline models.

Reference | Related Articles | Metrics

Select

Review of Speech Recognition Techniques for Low Data Resources

XU Chundong, WU Ziyu, GE Fengpei

Computer Engineering and Applications 2025, 61 (4): 59-71. DOI: 10.3778/j.issn.1002-8331.2405-0425

Abstract （54）

PDF（pc）（953KB）（61）

Save

Recently, the focus of automatic speech recognition has shifted from traditional methods to speech recognition methods based on deep learning. Moreover, the “large model” phenomenon reflects that the performance of deep learning methods significantly improves as the volume of training data increases. However, real-world complexity, uneven speech data distribution, and privacy concerns challenge data collection. Additionally, the annotation of speech data requires the involvement of a large number of professionals, leading to high labeling costs. Therefore, speech recognition often faces the issue of insufficient data resources in practical applications. Building a high-performing and stable speech recognition system under low data resource conditions remains a research challenge. Consequently, this paper briefly summarizes the development history of speech recognition, then outlines the basic framework of speech recognition and common open-source datasets at home and abroad. Focusing on the low data resource issue, this paper analyzes the methods for determining low data resources in detail, and then reviews four categories of technical solutions, including data augmentation, federated learning, self-supervised learning, and meta-learning, provides a systematic analysis of their performance status and advantages and disadvantages. Finally, this paper discusses the potential future development trends and possible challenges faced by this research direction.

Reference | Related Articles | Metrics

Select

DySnake-YOLO: Improved Detection of Surface Defects on YOLOv9c Circuit Board

LI Yaolong, CHEN Xiaolin, LIN Hao, WANG Yu, WANG Chunlin

Computer Engineering and Applications 2025, 61 (3): 242-252. DOI: 10.3778/j.issn.1002-8331.2408-0162

Abstract （82）

PDF（pc）（17567KB）（94）

Save

For the production of printed circuit boards with defects such as missing holes, open circuits, short circuits, burrs and false copper, and the low detection accuracy caused by problems such as the tiny size of the defects and the similarity of the background, this paper proposes a circuit board surface defect detection algorithm, DySnake-YOLO, that improves on YOLOv9. In the feature extraction part, a dynamic, query-aware sparse-attention mechanism, BRA, is added to perform fine-grained extraction of printed circuit board features. In the feature fusion section, a convolutional module RE4DConv is designed to fit the tubular scenario according to the tubular target to fit the board characteristics as well as to focus on the regional connectivity features, which enhances the ability of model to fuse the tubular-scale features in printed circuit boards. Experiments on the publicly available PCB defect dataset from Peking University show that the improved algorithm improves the mAP50 by 0.023 compared to the prototype, and the improved method improves the mAP50 and mAP50-95 by 0.071 and 0.085, respectively, compared to the mainstream target detection algorithms, such as YOLOv8n, which has a high value in the application of the printed circuit board defect detection task. The improved method has a high value in the task of printed circuit board defect detection.

Reference | Related Articles | Metrics

Select

Review on Optimization Algorithms for One-Stage Metal Surface Defect Detection in Deep Learning

DONG Jiadong, GUO Qinghu, CHEN Lin, SANG Feihu

Computer Engineering and Applications 2025, 61 (4): 72-89. DOI: 10.3778/j.issn.1002-8331.2408-0098

Abstract （58）

PDF（pc）（1716KB）（60）

Save

Scratches, pits, ripples and other defects on the metal surface will directly affect the quality of the product. Traditional detection methods are time consuming, and the accuracy is limited by the operator’s experience and skills. In recent years, breakthroughs of deep learning technology in the field of image recognition have provided new solutions for metal surface defect detection, and the deep learning-based metal surface defect detection method have achieved remarkable results in terms of detection accuracy and speed. In order to facilitate the research of metal surface defect detection algorithm, the optimization method and application of one-stage deep learning algorithm in metal surface defect detection are comprehensively analyzed. The commonly used metal surface defect datasets and algorithm evaluation indexes are introduced. The development history of object detection algorithms, the basic concepts and typical models of one-stage object detection algorithms are summarized. From three aspects of data enhancement, feature extraction and fusion, anchor frame optimization, the advantages and disadvantages of different algorithms and different optimization methods are compared and summarized, and the light weight of metal surface defect detection algorithm is also studied. The future research direction of metal surface defect detection algorithm is prospected from three aspects：multi-mode fusion, big data application technology, reality and virtual combination.

Reference | Related Articles | Metrics

Select

Robot Global Path Planning Based on Improved RRT Algorithm

CUI Xijie, WANG Xiaojun, LI Xiaohang

Computer Engineering and Applications 2025, 61 (4): 331-338. DOI: 10.3778/j.issn.1002-8331.2310-0006

Abstract （49）

PDF（pc）（1571KB）（59）

Save

Aiming at the problems of the RRT (rapidly exploring random tree) path planning algorithm, such as large search range, poor goal guidance, local minimum and tortuous path, an improved RRT path planning algorithm with limited adaptive sampling area is proposed. The search space is divided into uniform levels, and the sampling area is dynamically adjusted according to the level of new node and the number of sampling points in the level, so as to reduce the search range. The new node improvement strategy is used to make the random tree adjust to the target point adaptively according to the environment information and change the expansion step to generate new nodes. Then the obstacle avoidance strategy is used to increase the target orientation and obstacle avoidance performance of the algorithm. The improved inverse optimization and cubic B-spline curve which inserts nodes and reduces the turning angle are used to optimize the path. Compared with RRT algorithm, the search time and iteration times of this algorithm are reduced by more than 70% in different path environments, and the optimized path is shorter and smoother.

Reference | Related Articles | Metrics

Select

GCW-YOLOv8n: Lightweight Safety Helmet Wearing Detection Algorithm

XU Zhuang, QIAN Yurong, YAN Feng

Computer Engineering and Applications 2025, 61 (3): 144-154. DOI: 10.3778/j.issn.1002-8331.2409-0334

Abstract （85）

PDF（pc）（1750KB）（84）

Save

China is a major industrial country in the world. In various construction environments, the falling of construction materials and collisions on construction sites are the main causes of casualties. Accidents caused by head injuries often occur, and wearing safety helmets can ensure the safety of construction personnel to the greatest extent possible. In order to solve the problems of poor timeliness and low management efficiency of manual management, the existing models have strict requirements for computing power, large memory requirements, and handling of load and data transmission delay of industrial equipment, and to achieve edge computing and real-time control, a modified helmet wearing detection algorithm based on YOLOv8n is proposed. Firstly, a new GS-C2f module is proposed, which introduces GhostConv and SE (squeeze and excitation) attention mechanism, effectively reducing the computational complexity of the model and helping the network extract features effectively. Secondly, the CBAM attention mechanism is introduced in the Neck section to enhance the model focus on effective features. Finally, Wise-IoUv3 is introduced to further improve the accuracy of the model. Through experiments, compared with the original YOLOv8n model, this model achieves a 21.24% reduction in computational parameters and a 0.01 improvement in recognition accuracy, achieving satisfactory results between model accuracy and complexity.

Reference | Related Articles | Metrics

Select

Study on Application of Large Language Model in Constructing Knowledge Graph of Medical Cases of Rhinitis

LI Yue, HONG Hailan, LI Wenlin, YANG Tao

Computer Engineering and Applications 2025, 61 (4): 167-175. DOI: 10.3778/j.issn.1002-8331.2403-0379

Abstract （50）

PDF（pc）（5480KB）（56）

Save

An automated knowledge extraction method based on large language model is explored, aiming to construct a knowledge graph on the treatment of rhinitis by national medical master Gan Zuwang, and to provide innovative ideas and methods for the intelligent advancement in the field of traditional Chinese medicine. The clinical medical case data of professor Gan Zuwang are used as the base sample, and the ontology model is constructed using OWL (Web ontology language) to determine the extraction objects and relations, and then the prompt template combining the demonstration case and the relation list is used to guide the automated extraction experiments of the medical case data with the large language model, and the Nebula Graph is used for the storage and the visual display of the knowledge graph. Compared with the traditional knowledge extraction model Bert-BiLSTM-CRF, the ChatGPT4 model performs the best in terms of comprehensive indexes, with an F1 value of 82.75%, which provides an effective solution for the rapid processing of unstructured medical case data and achieves semi-automatic construction of knowledge graph in the field of Chinese medicine. The use of large language models for knowledge graph construction not only provides a practical solution for the intelligence in the field of Chinese medicine, but also contributes new research ideas for the inheritance of diagnostic and treatment experience of famous and veteran Chinese medicine practitioners and the rapid construction of the knowledge graph of Chinese medicine, which promotes the development of Chinese medicine.

Reference | Related Articles | Metrics

Select

MLDAC：Multi-Task Dense Attention Computation Self-Supervised Few-Shot Semantic Segmentation Method

WANG Weihang, ZHANG Yi

Computer Engineering and Applications 2025, 61 (4): 211-221. DOI: 10.3778/j.issn.1002-8331.2309-0439

Abstract （51）

PDF（pc）（1380KB）（55）

Save

Aiming at the problem that existing few-shot semantic segmentation methods still need a large number of pixel-level annotations to complete the training of models, a multi-task dense attention computation self-supervised few-shot semantic segmentation method (MLDAC) is proposed. The method divides the saliency of a single image in the dataset into two parts, one part serves as the support image mask for few-shot segmentation, the other part or the all saliency respectively makes the cross-entropy loss of the prediction result as multiple targets for multi-task learning, improving model generalization. The Swin Transformer is used for the backbone network to extract feature maps at different levels. These feature maps are input into multiple levels of dense attention computation blocks to enhance pixel-level correspondence. The final prediction results are obtained by using the inter-scale mixing and feature skip-connection. The experimental results indicate that MLDAC attains 55.1% and 26.8% 1-shot mIoU self-supervised few-shot segmentation on the PASCAL-5i and COCO-20i datasets respectively, compared with the current best self-supervised few-shot semantic segmentation method, improves by 1.3 and 2.2 percentage points respectively. In addition, the model achieves 78.1% 1-shot mIoU on FSS-1000 dataset, verifying its efficacy.

Reference | Related Articles | Metrics

Select

Survey on Lane Line Detection Techniques for Classifying Semantic Information Processing Modalities

HONG Shuying, ZHANG Donglin

Computer Engineering and Applications 2025, 61 (5): 1-17. DOI: 10.3778/j.issn.1002-8331.2406-0160

Abstract （56）

PDF（pc）（2981KB）（53）

Save

With the rapid development of autonomous driving technology, lane line detection, as its key component, has attracted widespread attention and shown great potential for application in intelligent transportation systems. However, traditional lane line detection techniques usually struggle to provide satisfactory recognition accuracy when dealing with complex environmental challenges. This paper reviews the development of lane detection technology and systematically sorts out 84 advanced algorithms, and innovatively divides them into four categories based on semantic processing: semantic segmentation assistance, semantic information fusion, semantic information enhancement, and semantic relationship mode-
ling. By deeply analyzing the technical characteristics and advantages of these algorithms, the main limitations of current lane line detection technology are revealed. Finally, the future development direction of lane line detection technology is put forward, especially in the utilization of semantic information, and the potential research direction is pointed out.

Reference | Related Articles | Metrics

Select

Improved YOLOv8s Model for Small Object Detection from Perspective of Drones

PAN Wei, WEI Chao, QIAN Chunyu, YANG Zhe

Computer Engineering and Applications 2024, 60 (9): 142-150. DOI: 10.3778/j.issn.1002-8331.2312-0043

Abstract （662）

PDF（pc）（5858KB）（802）

Save

Facing with the problems of small and densely distributed image targets, uneven class distribution, and model size limitation of hardware conditions, object detection from the perspective of drones has less precise results. A new improved model based on YOLOv8s with multiple attention mechanisms is proposed. To solve the problem of shared attention weight parameters in receptive field features and enhance feature extraction ability, receptive field attention convolution and CBAM (concentration based attention module) attention mechanism are introduced into the backbone, adding attention weight in channel and spatial dimensions. By introducing large separable kernel attention into feature pyramid pooling layers, information fusion between different levels of features is increased. The feature layers with rich semantic information of small targets are added to improve the neck structure. The inner-IoU loss function is used to improve the MPDIoU (minimum point distance based IoU) function and the inner-MPDIoU instead of the original loss function is used to enhance the learning ability for difficult samples. The experimental results show that the improved YOLOv8s model has improved mAP, P, and R by 16.1%, 9.3%, and 14.9% respectively on the VisDrone dataset, surpassing YOLOv8m in performance and can be effectively applied to unmanned aerial vehicle visual detection tasks.

Reference | Related Articles | Metrics

Select

Construction and Study of Explainable Logical Reasoning Dataset

XIAO Yu, XIAO Jing, LIN Guijin, NI Rongsen, XIAN Jiarong, YUAN Jibao

Computer Engineering and Applications 2025, 61 (4): 114-121. DOI: 10.3778/j.issn.1002-8331.2309-0458

Abstract （38）

PDF（pc）（937KB）（51）

Save

Logical reasoning ability is crucial to understand natural language by machines and humans. The explanation of logical reasoning problems is an elaboration and description of the logical reasoning process. However, such explanations are lacking in current logical reasoning benchmarks. For this problem, this paper creates a Chinese and English dataset called explainable logical reasoning (Ex-LoR). This dataset contains 3411 logical reasoning problems with explanation data, and categorizes these problems into 6 classes according to their reasoning methods. This paper designs two tasks：logical reasoning question and answer task, and explanation generation task. Subsequently, this paper conducts experiments and analysis on this dataset by using several language models. The results show that the existing language models are still unable to well answer logical reasoning questions and generate reasonable explanations. Therefore, it is challenging to equip machines with logical reasoning capabilities. The logical reasoning dataset and experimental results presented in this paper can be used as a benchmark for subsequent research.

Reference | Related Articles | Metrics

Select

Incorporating Pre-Trained Model and Attention Mechanism for Event Extraction

XIAO Lizhong, YIN Chenxu

Computer Engineering and Applications 2025, 61 (4): 130-140. DOI: 10.3778/j.issn.1002-8331.2309-0255

Abstract （46）

PDF（pc）（999KB）（51）

Save

Event extraction aims to extract structured event information from a large amount of unstructured texts, but existing research work has problems such as difficulty in extracting overlapping roles, lack of interaction between subtasks, and insufficient semantic feature expression ability. This paper proposes a Chinese event extraction model PACJEE (pre-trained language model and attention mechanism based Chinese joint event extraction) to address these problems. The model uses the pre-trained language model RoBERTa to extract text features, then classifies the event types of the text, in the trigger word recognition stage, fuses the extracted event type prior features with the text features, and uses self-attention mechanism to obtain the internal feature relevance, in the argument role classification stage, introduces convolutional neural network and attention mechanism to enhance the trigger word feature expression ability, and finally uses multi-layer pointer tagging to identify overlapping roles. The method is experimentally analyzed on the Chinese datasets ACE2005 and DuEE, and the results show that compared with the baseline methods, the F1 values of trigger word classification are increased by 1.6 and 0.5 percentage points respectively, and the F1 values of argument role classification are increased by 3.3 and 2.5 percentage points respectively, indicating that the model can significantly improve the event extraction effect, and to a certain extent, improve the recognition accuracy of overlapping role events.

Reference | Related Articles | Metrics

Select

Traditional Chinese Medicine Question Answering Model Based on Chain-of-Thought and Knowledge Graph

YUAN Zhongxu, LI Li, HE Fan, YANG Xiu, HAN Dongxuan

Computer Engineering and Applications 2025, 61 (4): 158-166. DOI: 10.3778/j.issn.1002-8331.2408-0101

Abstract （49）

PDF（pc）（1003KB）（49）

Save

In response to the large scale of data in the field of Chinese medicine diagnosis, as well as the high subjectivity of doctors in diagnosis and the difficulty of data alignment, ChatTCM, a large language model for the Q&A domain of traditional Chinese medicine (TCM), is proposed. Taking advantage of the power of large language model (LLM) in dealing with natural language understanding and text generation, and fine-tuning the large language model, the LLM has expertise and competence in the field of TCM Q&A, thus preventing the model from generating hallucinations. Firstly, extract triplet information from TCM books to construct a TCM knowledge graph database, achieving data alignment and systematic integration of TCM knowledge, while providing background knowledge for large language models to generate answers. Secondly, integrate chain-of-thought (COT) reasoning with dynamic interactions from the knowledge graph database to generate an objective reasoning process, ensuring that the diagnostic recommendations are based on scientific evidence. Additionally, store the reasoning results from the chain-of-thought and knowledge graph as new knowledge, continuously expanding the local knowledge base. The ChatTCM model improves the BLEU-4 and ROUGE-L metrics on the MedChatZH dataset by 10.6 and 10.5 percentage points, respectively, and achieves 70% accuracy on the open-source dataset, which is a 10 percentage points improvement over the same type of MedChatZH model.

Reference | Related Articles | Metrics

Select

Review on Improvement and Application of 3D Convolutional Neural Networks

LI Zehui, ZHANG Lin, SHAN Xianying

Computer Engineering and Applications 2025, 61 (3): 48-61. DOI: 10.3778/j.issn.1002-8331.2407-0031

Abstract （60）

PDF（pc）（890KB）（72）

Save

3D convolutional neural network, as a kind of deep neural network, has shown excellent results in the field of computer vision, especially in video action recognition. However, there are still some problems in 3D convolutional neural networks. In order to solve these problems, this paper summarizes and analyzes the existing improved methods of video action recognition based on 3D convolution. The improvement of 3D convolutional neural network is summarized in the aspects of lightweight, feature extraction, computational efficiency, combination model, etc. The practical application of 3D convolutional neural network is introduced, the popular data sets are summarized, and the experimental results of these improved methods are compared and analyzed. Finally, the future development direction of video action recognition is prospected.

Reference | Related Articles | Metrics

Select

Simple and Effective Weakly Supervised Chinese Text Classification Algorithm

CHEN Zhongtao, ZHOU Yatong

Computer Engineering and Applications 2025, 61 (4): 192-210. DOI: 10.3778/j.issn.1002-8331.2310-0009

Abstract （42）

PDF（pc）（4169KB）（43）

Save

Most of the current weakly supervised text classification algorithms based on seed words need to search all seed words from the dataset and extend the category dictionary in this way, and the category recognition ability of seed words that occur less frequently is also lower. Therefore, a simple and effective weakly supervised Chinese text classification (SEWClass) algorithm is designed, which uses the initial weights of the pre-trained language model to generate an abstract understanding of the text and continues to generate abstract constraints and figurative constraints based on this to construct the initial training. Based on the number of categories, a dimensionality reduction model and a classifier are jointly constructed to adapt to the fact that the weakly supervised text classification needs to be specified in advance, and needs to increase training data during self-training. With the two constraints, the pseudo-labeled data have a high precision rate, and only the dimensionality reduction model is trained during self-training to improve the recall and efficiency. SEWClass requires only one seed word, such as the category name, to complete the classification task, and the performance of SEWClass is independent whether or not the seed word occurs in the dataset. The performance of SEWClass on both Chinese datasets, THUCNews and toutiao, is much higher than that of other weakly supervised algorithms.

Reference | Related Articles | Metrics

Select

Defect Detection of Photovoltaic Cells Based on RFCARep-YOLOv8n

ZHANG Ji, WANG Wenbin, YU Yang

Computer Engineering and Applications 2025, 61 (3): 131-143. DOI: 10.3778/j.issn.1002-8331.2405-0249

Abstract （52）

PDF（pc）（2043KB）（68）

Save

To address the problems of target occlusion, complex background and small target defects that are difficult for the human eye to distinguish in the defect images of photovoltaic cells, an photovoltaic cell defect detection algorithm with receptive field coordinated attention re-parameterization-YOLOv8n（RFCARep-YOLOv8n） is proposed. Firstly, a receptive-field coordinated attention re-parameterization block is proposed to replace the bottleneck block for feature extraction, expand the attention to global information to improve semantic expression ability, and suppress the interference of occlusion and complex background. Secondly, a large separable kernel attention module is added after spatial pyramid pooling-fast to enhance the global feature information fusion by improving the long-distance feature dependence. Finally, in the feature fusion part, multi-scale sequence fusion neck is used, and the multi-scale auxiliary detection head is combined to reduce the loss of detailed features and improve the detection ability of small target defects. Experimental results show that on the PASCAL VOC dataset the proposed model is 2.3 and 2.1 percentage points higher than the baseline model compared with mAP@0.5 and mAP@0.5:0.95, and on the photovoltaic defect dataset reaches 87.6% in mAP@0.5, which is 3.5 percentage points higher than the baseline model, and the parameter number is 3.23×106. The lightweight parameters of the benchmark model are maintained while the detection performance is improved.

Reference | Related Articles | Metrics

Select

Review of Human Action Recognition Based on Skeletal Graph Neural Networks

JIANG Yuehan, CHEN Junjie, LI Hongjun

Computer Engineering and Applications 2025, 61 (3): 34-47. DOI: 10.3778/j.issn.1002-8331.2407-0391

Abstract （71）

PDF（pc）（14282KB）（75）

Save

Human action recognition based on skeletal graph neural network has attracted much attention by virtue of its simplicity and robustness, and graph data have a natural advantage in processing human skeletal information, which has gradually become a research hotspot in the field of action recognition. Starting from the broad basic concept of action recognition, this paper further introduces the task of human action recognition using skeletal graph neural networks, and summarizes the research results on human action recognition using skeletal graph neural networks in recent years from four aspects. The paper introduces different methods for constructing topological graphs with graph structures, the common mechanisms used in action recognition models, the comparison of commonly used datasets and evaluation indicators with mainstream methods. Finally, the problems of human action recognition based on skeletal graph neural networks are elaborated in detail with respect to the current state of research, and an outlook on the future development of the field is given based on the current state of research.

Reference | Related Articles | Metrics

Select

Small Target Detection Algorithm for UAV Based on Composite Feature and Multi-Scale Fusion

LIAO Ningsheng, CAO Tianxiu, LIU Keyan, XU Meng, ZHU Mi, GU Yuxuan, WANG Pengfei

Computer Engineering and Applications 2025, 61 (3): 111-120. DOI: 10.3778/j.issn.1002-8331.2407-0520

Abstract （63）

PDF（pc）（3642KB）（65）

Save

A UAV target detection algorithm with composite feature and multi-scale fusion, CM-YOLOv8s(composite and multi-scale YOLOv8s) is proposed to address the issues of missed detections, false detections, and the imbalance between accuracy and speed in current UAV perspective detection algorithms. The quality of target composite features is improved by introducing channel features in the spatial pyramid pooling module. The neck structure of the model is redesigned to improve the retention ratio of detailed target features. Additionally, the DRHead detection head is designed to achieve multi-scale feature map fusion, enhancing the adaptability for multi-scale target detection. The Wise-IoU loss function is employed to accelerate model convergence. Compared to the baseline algorithm, the improved CM-YOLOv8s algorithm has a parameter count of only 3.5×106, reducing the parameters by 69%. Experimental results show that the proposed CM-YOLOv8s algorithm significantly improves the mAP50 by 6.8 percentage poins on the VisDrone2019 dataset. Furthermore, the generalization and effectiveness of the proposed algorithm are validated on the UAV-DT and DIOR datasets.

Reference | Related Articles | Metrics

Select

Improved YOLOv8s Model for Smoke and Flame Detection in Complex Backgrounds

MA Yaoming, ZHANG Pengfei, TAN Fusheng

Computer Engineering and Applications 2025, 61 (3): 121-130. DOI: 10.3778/j.issn.1002-8331.2406-0348

Abstract （68）

PDF（pc）（1564KB）（64）

Save

Aiming to address issues such as confusion between smoke flame targets and background within complex backgrounds, which often result in low accuracy of smoke flame detection, an enhanced model based on YOLOv8s for detecting smoke flames within complex backgrounds is proposed. Firstly, the feature channels are highly similar to each other, and in order to effectively utilize the redundancy across different channels and improve the ability of model to differentiate between smoke and flame targets and backgrounds, the C2fFR (C2f with partial rep conv) lightweight feature extraction module is introduced. Secondly, the MCFM (multi-scale context fusion module) is designed to capture and utilize contextual information for enhancing feature representation. Lastly, the Inner-SIoU loss function is employed to address bounding box mismatches and the regression ability of the model is improved for high IoU samples. Experimental results demonstrate that compared to the baseline YOLOv8s model, the enhanced YOLOv8s smoke flame detection model achieves improvements of 4.6 percentage points in mAP@50 and 2.3 percentage points in mAP@50:95. Moreover, it reduces the number of model parameters by 18.9% and computation by 8.1%. while maintaining an FPS (frame per second) of 93. Additionally, it exhibits superior detection performance when compared to other mainstream detection algorithms.

Reference | Related Articles | Metrics

Select

Semi-Supervised Medical Image Segmentation with Label-Part Switching and Cross-Teaching

LUO Yiheng, ZHANG Junhua, ZHANG Jianqing

Computer Engineering and Applications 2025, 61 (4): 253-261. DOI: 10.3778/j.issn.1002-8331.2310-0062

Abstract （43）

PDF（pc）（2548KB）（39）

Save

A semi-supervised medical image segmentation algorithm incorporating label-part switching and cross-teaching is proposed in response to the current challenges in the field of medical image segmentation, including low segmentation accuracy and high costs and difficulty in data acquisition. The label-part switching algorithm locates and exchanges the label portions of two images, addressing issues related to data distribution mismatch and empirical distribution gaps. The Transformer network is applied in the Mean Teachers architecture, employed for cross-teaching with CNN to assist in improving the quality of pseudo-label generation. A training strategy is introduced for images with swapped labels during pre-training and self-training, expanding the training dataset to enable the model to learn more features. In experiments with 10% labeled data on the ACDC dataset, the Dice coefficient reaches 90.67%, showing a 2.26 percentage points improvement over the baseline model. In experiments with 5% labeled data on the ACDC dataset and 20% labeled data on the PROMISE12 dataset, the Dice coefficients reach 88.69% and 84.34%, respectively. Comparative experiments with other methods demonstrate optimal performance across various metrics, validating the effectiveness and reliability of the proposed approach.

Reference | Related Articles | Metrics

Select

Two-Stage Feature Transfer Image Dehazing Algorithm

YUAN Heng, YAN Tinghao, ZHANG Shengchong

Computer Engineering and Applications 2025, 61 (4): 241-252. DOI: 10.3778/j.issn.1002-8331.2309-0455

Abstract （38）

PDF（pc）（11350KB）（38）

Save

To solve the problems such as artifacts, color distortion and unsatisfactory dehazing effect on images under the influence of non-uniform fog after image processing by common dehazing algorithms, a two-stage feature transfer image dehazing algorithm is proposed, which is implemented based on the encoder-decoder structure. In the first stage, the clear image is sent to the feature learning network, and the spatial structure information and color rules of the clear image are learned through the hybrid attention mechanism. In the second stage, the feature transfer loss is used to transfer the clear image feature knowledge learned in the feature learning network to the feature refinement image dehazing network. At the same time, the image context information is effectively extracted and fused through the multi-scale feature extraction module and the global feature refinement block. Finally, the output of the two stages is fused to restore a clear and dehazing image. The experimental results show that the algorithm has a good dehazing effect in the RESIDE dataset and real non-uniform foggy images, and the color of the processed image is reasonable and more in line with human visual perception.

Reference | Related Articles | Metrics

Select

Adaptive Iterative Learning for Predicting Metrics of Dispatching Automation System

SHEN Jialing, JI Xuechun, GAO Shang, WANG Yudong, CHEN Ziyun, LI Hao

Computer Engineering and Applications 2025, 61 (4): 368-376. DOI: 10.3778/j.issn.1002-8331.2309-0462

Abstract （43）

PDF（pc）（989KB）（38）

Save

A adaptive iterative learning method for predicting metrics of dispatching automation systems is proposed to address issues such as low accuracy of a single algorithm applied to massive metrics and failure to iteratively update based on real-time data feature changes in the context of intelligent risk warning scenarios. Based on the temporal data characteristics of metrics under different behavioral patterns of business applications in power system, a classification method for metrics based on Fourier transform and autocorrelation coefficient is proposed. Based on the classification results, an adaptive selection strategy is adopted to construct a timeseries prediction model for the metric. Real-time metric changes are dynamically captured and adaptively iterated to update model and prediction results. This paper selects some metric data of a system for example analysis to verify that the proposed method is significantly better than a single algorithm in terms of accuracy and timeliness, eliminating the impact of real-time data feature changes on metric prediction.

Reference | Related Articles | Metrics

Select

Computer Engineering and Applications 2025, 61 (3): 0-0.

Abstract （76）

PDF（pc）（690KB）（50）

Save

Most Download articles