基于改进YOLOv4-tiny的铝合金型材表面缺陷检测

Surface defect detection of aluminum alloy profiles based on modified YOLOv4-tiny

为了解决铝合金的擦花、喷流、脏点三类表面缺陷检测时,YOLOv4-tiny模型原有目标先验框对缺陷目标尺寸不敏感、缺陷检测精确度低等问题,利用K-means聚类算法将目标边框重新聚类,生成6个新的先验框,增强先验框在对目标检测时的敏感性,将通道注意力机制的典型代表SEnet添加到YOLOv4-tiny的加强特征提取网络中,提高检测精确度。结果表明,改进后的网络每秒传输帧数为94,与原网络相比均值平均精度提高3.46%,精确度和召回率的调和平均值提高2.67%。改进后网络占用内存为22.6 MB,能够适应工业环境下实时检测的精准性。

This paper is focus on a study in response to the insensitivity to the size of defect target and low accuracy in defect detection with the original target priori box of YOLOv4-tiny model,as occurs in the process of detecting surface defects in aluminum alloy including scratches,jet and dirty spots.The study involves using K-means clustering algorithm to re-cluster the target border and generate six new priori boxes for enhancing the sensitivity of the priori box in target detection;adding SEnet,a typical representative of channel attention mechanism,to the enhanced feature extraction network of YOLOv4-tiny to improve detection accuracy.The results show that the improved network has 94 frames per second;the mean average precision increases by 3.46%,the harmonic mean mixed by accuracy and recall rate increases by 2.67%.The improved network occupies the memory by 22.6 MB,which is suitable for the efficient and accurate real-time detection in industrial environment.