Surface Defect Detection Method for Steel Bar Based on Correction Attention YOLOF

In order to accurately detect the surface defect of steel bar in the rolling process in real time, a surface defect detection method for steel bar based on correction attention YOLOF is proposed. Taking the YOLOF as the backbone network in this method. And a correction attention network branch fused with YOLOF is designed, which uses the gray correction module and outlier segmentation module to construct the input and output channel for parameter learning, and the intermediate features generated by the attention network module are fused with the features of the YOLOF, which are used to enhance the contextual semantic information. The results show that compared with the original YOLOF, the mean average precision (mAP) of the proposed method is improved by 7.69%. In addition, in order to verify that the correction attention network branch has generalization ability and improve the detection effect, the branch is added to TinyYOLOv3 and SSD512 respectively, increasing the mAP by 6.85% and 4.98% respectively.