关键词: 目标检测, 样本选择策略, 宽高比, 金属缺陷检测, 距离回归损失函数

Abstract: The distance-aware dynamic label assignment(DDA) algorithm was proposed to address issues such as the lack of consideration for aspect ratio of metal surface defects and poor localization ability towards target distribution during the allocation of positive and negative samples in training processes of metal surface defect detection model. DDA did not change the structures of original detection model and did not increase computational expenses. A new distance loss calculation paradigm was proposed based on geometric characteristics of real frame to optimize the regression problem with a wide aspect ratio. The regression offset in iterative processes was decoded as predicted frame coordinates. Finally, the comprehensive intersection and union ratio information were calculated among the predicted frame, anchor frame and real frame, and positive and negative samples were dynamically selected to improve training accuracy. It was verified through the surface defect detection task of cold-rolled strip in a steel plant in Wuhan, and a public hot-rolled strip surface defect data set was introduced for generalization testing. The detection results are significantly improved, which has practical application values for metal surface quality specifications.

Key words: object detection, sample selection strategy, aspect ratio, metal defect detection, distance regression loss function