SWP-YOLO: efficient pitting defect detection network for ball screws

Aiming at the problem of small and difficult to detect pitting defects in ball screws, an improved YOLO11 pitting detection model for ball screws is proposed. This model references the space-to-depth convolution module, which improves the detection performance of the model for small targets. It is also equipped with a special inspection head for small defect detection, constituting a four-head structure. Secondly, the innovative design of C3k2_DWR module and Detect_PSP module replaces the original C3k2 module with C3k2_DWR module, which can accurately extract the pitting features, especially for small target detection. The results indicate that the mAP@50 value of the SWP-YOLO model has reached 89.6%, which is 5% higher than that of YOLO11n, and its model size is also only 5 MB and 6.2 GFLOPs of computational complexity. The inference speed reaches 318 FPS, achieving an excellent balance between accuracy, lightweight design and real-time performance. SWP-YOLO can quickly and accurately identify the pitting defects of ball screws, providing an efficient new solution for industrial visual inspection.