摘要
针对YOLOv5对金属表面细小缺陷或微观缺陷的检测结果易受背景干扰的问题,提出了一种改进的金属表面缺陷检测算法。通过在主干网络引入坐标注意力机制,提高模型对缺陷的关注度;将主干网络中的一些CBS和C3模块替换为GhostNetV2结构构建轻量级的网络,优化模型的性能和效率;在Neck层采用双向特征融合网络(BiFPN)来增强颈部以产生丰富的表征,加深整个网络并重用低层次的特征。最后,广泛的实验结果表明,CGB-YOLO在NEU-DET上的精度达到75.0%mAP,比改进前提高了3.8%。该模型在金属表面缺陷检测中具有较好的综合性能。
In order to solve the problems of YOLOv5 on the problems of too many small targets on metal surface defects and the detection results are easy to be interfered by background,an improved metal surface defect detection algorithm was proposed.By introducing the coordinate attention mechanism in the backbone network,the model pays attention to defects,and some CBS and C3 modules in the backbone network were replaced with GhostNetV2 structure to build a lightweight network to optimize the performance and efficiency of the model.A bidirectional feature fusion network(BiFPN)was used to enhance the neck layer to generate rich representations,deepen the whole network and reuse low-level features.Finally,extensive experimental results showed that the accuracy of CGB-YOLO on NEU-DET reaches 75.0%mAP,which is 3.8%higher than that before the improvement.The model has good comprehensive performance in metal surface defect detection.
作者
任洛莹
刘兴德
谢延楠
胡文松
余鹏泽
孔志成
REN Luoying;LIU Xingde;XIE Yannan;HU Wensong;YU Pengze;KONG Zhicheng(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin City 132022,China;School of Mechanical and Electrical Engineering,Jilin Institute of Chemical Technology,Jilin City 132022,China)
出处
《吉林化工学院学报》
CAS
2023年第11期38-44,共7页
Journal of Jilin Institute of Chemical Technology
基金
吉林省科技厅项目(20190302080GX)
吉林市科学技术局项目(201750209)
作者简介
通信作者:刘兴德,E-mail:396803006@qq.com;任洛莹,吉林化工学院2022级学生;谢延楠,吉林化工学院2022级学生;胡文松,吉林化工学院2022级学生;余鹏泽,吉林化工学院2022级学生;孔志成,吉林化工学院2022级学生
