引用本文: | 杨 乐,郭一鸣,霍勇博,等.改进YOLOv5在电力生产违规穿戴检测中的应用[J].电力系统保护与控制,2023,51(14):160-168.[点击复制] |
YANG Le,GUO Yiming,HUO Yongbo,et al.Application of improved YOLOv5 for illegal wearing detection in electric power construction[J].Power System Protection and Control,2023,51(14):160-168[点击复制] |
|
摘要: |
为了解决现有目标检测系统在电力现场识别中存在的环境复杂、检测物体形状方差过大以及视觉特征辨识性不佳等问题,提出了一种适用于电力现场穿戴识别的目标检测模型。首先,通过在YOLOv5特征提取网络中嵌入非对称卷积模块,从而得到更加具备辨识性及鲁棒性的视觉特征。其次,为了能够在全局背景噪声的影响下自适应地关注与检测物体特征相关性更强的区域,采用全局注意力机制进行上下文信息的建模,改进了视觉信息处理的效率与准确性。最后,通过对比现有的目标检测算法,证明了所提针对YOLOv5改进算法的有效性和优越性。同时,通过消融实验证明了所改进的模块在目标检测模型中的有效性。 |
关键词: 非对称卷积网络 注意力机制 目标检测 违规穿戴检测 YOLOv5 |
DOI:10.19783/j.cnki.pspc.221785 |
投稿时间:2022-11-09修订日期:2023-02-09 |
基金项目:国家自然科学基金资助项目资助(62176227,U2066213);中央高校基本科研业务费资助(20720210047);国家电网陕西省电力公司科技项目资助(SGSNXT00GCJS2200106) |
|
Application of improved YOLOv5 for illegal wearing detection in electric power construction |
YANG Le1,GUO Yiming1,HUO Yongbo1,REN Xiaolong1,LIN Pingyuan2,ZHANG Zhihong2 |
(1. State Grid Shaanxi Electric Power Company Information and Communication Co., Ltd., Xi'an 710065, China;
2. School of Informatics, Xiamen University, Xiamen 361000, China) |
Abstract: |
There are problems in existing object detection systems caused by the complicated detection environment of electric power construction, large variance of detected object shape and poor recognition of visual features. This paper proposes an object detection model for illegal wearing detection in electric power construction. First, an asymmetric convolutional group is added to the feature extraction backbone network of YOLOv5 to obtain more discriminative and robust visual features. Second, in order to be able to adaptively focus on feature regions that are more relevant to the detected object under the influence of visual noise, a transformer-based global attention mechanism for modeling contextual information is used to improve the efficiency and accuracy of visual information processing. Finally, the effectiveness and superiority of the improved YOLOv5 algorithm in this paper are demonstrated by comparing with existing object detection algorithms. Meanwhile, the effectiveness of the improved modules in the target detection model is demonstrated by ablation experiments. |
Key words: asymmetric convolutional network attention mechanism object detection illegal wearing detection YOLOv5 |