Abstract:In view of the problems of insufficient detection accuracy and high missed detection rate in bearing surface defect detection, a defect detection model DSRC-YOLOv11n based on improved YOLOv11n is proposed. Firstly, a lightweight backbone network DS-Backbone is designed: Split Spatial Compression Convolution (SPConv) is embedded to reduce the computational cost of the backbone network, and Dynamic Convolution (DynamicConv) is introduced into the core C3k2 module to enhance the adaptability to complex defects. Secondly, the Dilation-wise Residual (DWR) module is introduced into the neck network to construct C3k2-DWR, expanding the receptive field to strengthen multi-scale feature fusion. Finally, the Channel Prior Convolutional Attention mechanism (CPCA) is introduced to guide the network to focus on key defect regions. Experiments on the self-built bearing surface defect dataset show that compared with the original YOLOv11n model, the recall rate, mAP50, and mAP @0.5:0.95 of DSRC-YOLOv11n are increased by 4.2%, 3.0%, and 2.1% respectively. The significant improvement in the recall rate effectively reduces the missed detection rate. This model effectively controls missed detections while improving detection accuracy, providing reliable technical support for automated quality inspection in industrial scenarios.