Surface Defect Detection of Steel Strips Based on Anchor-Free Network With Channel Attention and Bidirectional Feature Fusion

Abstract

Strip steel is an indispensable material in the manufacturing industry and the defects of the surface directly determine the quality. Due to the diversity and complexity of surface defects in intraclass and between interclass, a great deal of manpower and resources have been devoted to surface defect detection. This article proposes a new deep learning detection network, channel attention, and bidirectional feature fusion on a fully convolutional one-stage (CABF-FCOS) network to achieve rapid and effective defect detection on steel strips. First, the anchor-free FCOS is proposed as the detection framework to eliminate affections of those hyperparameters related to the anchor. Second, a channel attention mechanism (CAM) module is proposed to reduce the loss of feature information. Finally, it replaces the feature pyramid network (FPN) with a bidirectional feature fusion network (BFFN) for more effective feature fusion on the images. This reduces the loss of feature information and performs feature fusion to improve the defect detection performance. The experiments show that the mean average precision (mAP) is 76.68% at a detection speed of 18 frames per second (FPS) and improved by 4.43% over FCOS, which is higher than the state-of-the-art (SOTA) detection methods. This indicates that CABF-FCOS can obtain satisfactory defect detection performance and meet the demand of real-time detection in the industry.