RDD-YOLO: A modified YOLO for detection of steel surface defects

Abstract

Steel surfaces may exist some defects owing to imperfect manufacturing crafts and external factors, which seriously influence the lifespan and availability of steel. Thus, surface defect detection is a necessary process during industrial production. However, traditional surface defect detection algorithms have the shortcomings of low accuracy and speed. Therefore, we propose a model, named RDD-YOLO, based on YOLOv5 for steel surface defect detection. Firstly, the backbone component is consisted of Res2Net blocks to enlarge the receptive field and extract features of various scales. Secondly, in the neck, we design a double feature pyramid network (DFPN) to enhance the neck to generate abundant representations, which deepens the whole network and reuses the features of low levels. Thirdly, a decoupled head is employed to separate the regression and classification tasks for more detection precision. Finally, the extensive experimental results illustrate that the accuracies of RDD-YOLO reach 81.1 mAP on NEU-DET and 75.2 mAP on GC10-DET and improve by 4.3% and 5.8% over YOLOv5, respectively. Our proposed model achieves a comprehensive performance in steel surface defect detection.