基于深度学习的曲面玻璃表面缺陷检测方法

针对曲面玻璃表面缺陷成像难、识别准确率低等问题，提出一种基于YOLOv4的曲面玻璃表面缺陷检测方法。根据光源的方向确定平面与曲面的光学特性，采用明场背面漫射照明的方式来获得图像信息，确立打光方案后获取不同表面的缺陷图片。使用改进K-means聚类算法，采用交并比函数确定锚框的量度，解决原锚框大小不适用于玻璃缺陷小目标检测问题。将所提方法与缺陷检测主流算法对比验证。结果表明：所提改进的YOLOv4方法均值平均精度(mAP)可以达到80.14%,与Faster RCNN以及YOLOv3算法相比，mAP分别提升了8.29%和16.11%,并且有更好的鲁棒性和检测效果。

A Detection Method for Surface Defects of Curved Glass Based on Deep Learning

Aiming at the problems of difficulty in imaging the surface defects of curved glass and low recognition accuracy,a detection method of curved glass surface defects based on YOLOv4 was proposed.According to the direction of the light source,the optical characteristics of the plane and the curved surface were determined,and the bright-field backside diffused illumination was used to obtain image information.After the lighting plan was established,the defect pictures of different surfaces were obtained.The improved K-means clustering algorithm was used,and the function of the intersection over union was used to determine the measurement of the Anchor frame,which solved the problem that the size of the original Anchor frame was not suitable for the detection of small object with glass defects.The proposed method was compared and verified with mainstream defect detection algorithms such as Faster R-CNN and YOLOv3.The results show that the mean average precision (mAP) of the improved YOLOv4 method can reach 80.14%,the mAP is improved by 8.29% and 16.11%,and it has better robustness and detection effect.