基于YOLOv4算法的冲压件缺陷检测

Defect detection of stamping parts based on YOLOv4 algorithm

针对冲压件缺陷检测目前存在的人工检测强度大、效率低等问题,提出了一种基于改进YOLOv4(You Only Look Once)模型的快速检测算法(YOLOv4-Mobile)。该方法使用改进的MobileNetV3网络代替YOLOv4结构中的CSPDarknet53网络,改进的MobileNetV3网络结合了深度可分离卷积、具有线性瓶颈的倒残差结构以及SE结构(轻量级注意力结构)。利用车间采集的冲压件图像,建立缺陷数据集并进行数据增强,使用K均值(K-means)聚类算法得到一组对应冲压件缺陷数据集的先验框参数,提高了先验框与特征图层的匹配度。实验结果表明:基于改进YOLOv4模型的快速检测算法的平均精度达到89%,高于SSD算法;同时,单张检测时间达到0.15 s,优于原有的YOLOv4算法。

For the problems of high manual detection intensity and low efficiency in defect detection of stamping parts at present, a fast detection algorithm(YOLOv4-Mobile) based on the improved YOLOv4(You Only Look Once) model was proposed, which used the improved MobileNetV3 network to replace the CSPDarknet53 network in YOLOv4 structure, and the improved MobileNetV3 network combined a depthwise separable convolution, an inverted residual structure with a linear bottleneck and SE(Squeeze and Excitation) structure. Then, the image of stamping parts collected in the workshop was used to establish the defect data set and enhance the data set, and a set of prior frame parameters corresponding to the defect data set of stamping parts was obtained by K-means clustering algorithm to improve the matching degree of prior frame and feature layer. The test results show that based on the improved YOLOv4 model, the mAP(mean Average Precision) of the fast detection algorithm reaches 89%, which is higher than that of SSD algorithm. Meanwhile, the detection speed reaches 0.15 s per sheet, which is better than the original YOLOv4 algorithm.