Application of Multi-Scale Feature Fusion and Deep Learning in Detection of Steel Strip Surface Defect

Abstract

Surface defect detection forms an integral part of steel quality control. There are many challenges in defect detection. Firstly, the sample size of steel strip is small which is leading to the over-fitting and low detection rate. Secondly, there are many tiny and slender defects on the surface of steel strip, and most of them have less information available for detection. In this paper, firstly, we augment the NEU dataset of six kinds of typical surface defects on hot-rolled steel strip to increase detection rate. Then, we improve Faster R-CNN and FPN networks and adopt multi-scale feature fusion to encourage the reusing of low-level features. Finally, we conducted detection experiments to evaluate the effect of data augmentation and multi-scale feature fusion. The result shows that the mAP of our network reached 98.26% and our algorithms achieved higher detection accuracy and located the position of defects more accurately, especially enhanced the detection effect of tiny and slender defects.