YOLO-CDP: An Improved Algorithm for Bearing Surface Defect Detection

To address the challenges of low detection accuracy for tiny defects and the issue of large model sizes in bearing surface defect detection, we present YOLO-CDP, a lightweight detection algorithm developed on the foundation of YOLOv8n. First, we design the C2f_DSConv module to enable the model to sense defects of different scales and shapes, and improve the accuracy of defect identification and localization. Second, we use the feature-based content-aware reassembly upsampling operator CARAFE in the neck, helping the model capture richer contextual information and enhancing its feature representation capacity. Finally, we add a new small object detection layer for detecting small-size defects on the bearing surface. Experimental results indicate that YOLO-CDP achieves a mAP of 92.4%, an increase of 3.8% over the original algorithm. The model has only 3.4M parameters, ensuring lightweight efficiency while enhancing detection accuracy.