RDB-YOLO: A Lightweight and High-Precision Steel Surface Defects Detection Method

Most deep learning-based target detection algorithms have low speed and accuracy for detecting defects on steel surfaces, which fails to meet the requirements of industrial sites. To tackle this important issue, we propose a lightweight network structure based on YOLOv5n. Firstly, Receptive Field Enhancement (RFE) is introduced into the backbone network for compensation for loss of accuracy caused by light changes and environmental noise interference. Secondly, Deformable Convolution network (DCN) is utilized to improve performance of the network to fit defects of different shapes. Thirdly, Bi-Ievel Routing Attention mechanism (BRA) is embedded into the neck to improve detection performance in small targets. Experimental data from the NED -DET dataset indicates that the Mean Average Precision (mAP) of the improved algorithm is 82.1 %, which is 5.1 % higher than that of the benchmark model, and the detection rate is 60.4 frame/s. Compared with other mainstream object detection algorithms, our algorithm has higher accuracy, lighter weight parameters and can achieve efficient steel surface defect detection under the premise of maintaining a certain detection speed.