摘要
针对面板表面缺陷检测存在精度低、效率低等问题,提出了一种基于Faster R-CNN的优化缺陷检测的算法。该方法通过在特征融合层添加不降维的局部自适应跨通道卷积,以增加通道交叉的特征映射;且在骨干特征提取网络后加入CBAM注意力网络,从而捕获特征图的长期特征依赖关系。并分析了缺陷数据集样本宽高比的差异性,设定锚框生成大小,结合DIoU-NMS建议框筛选机制以提升先验框与目标框的匹配率。实验结果表明,优化后网络模型的精确率与识别率均得到很大提升。
In view of the low precision and low efficiency of panel surface defect detection,this paper proposes an optimized defect detection algorithm based on Faster R-CNN.This method adds local adaptive cross-channel convolution without dimensionality reduction in the feature fusion layer to increase the feature mapping of channel crossing,and adds the CBAM attention network after the backbone feature extraction network to capture the long-term feature dependency of the feature map.It also analyzes the difference in the aspect ratio of the defect data set,sets the generation size of the aiming window,and combines the DIoU-NMS suggestion frame screening mechanism to improve the matching rate of the prior frame and the target frame.Experimental results show that the accuracy and recognition rate of the optimized network model have been greatly improved.
作者
陈婉琴
唐清善
黄涛
Chen Wanqin;Tang Qingshan;Huang Tao(School of Physics and Electronic Science,Changsha University of Science and Technology,Changsha 410000,China;No.3303 Factory of PLA,Wuhan 430200,China)
出处
《电子技术应用》
2022年第1期133-137,共5页
Application of Electronic Technique
