ASW-YOLO: Lightweight Gear Defect Detection with Improved YOLOv8n.

Aiming at the problems of diverse defect types, large-scale differences, and complex background interference in gear surface defect detection, a lightweight model, ASW-YOLO, is proposed based on YOLOv8n. By using an ADown dual downsampling module to compress feature map resolution and preserve fine-grained information. C2f_SE channel attention is introduced to enhance small-scale defect response. The CIoU is replaced with WIoU to optimize multi-scale target localization accuracy. The experiments are conducted on the gear dataset. The comparative experiments show that mAP@0.5 of ASW-YOLO reached 94.8%, an increase of 4.5% compared to YOLOv8n, with a reduction of 9.3% in parameter count and 8.5% in computational complexity. The ablation experiments confirm the effectiveness of the three modules. ASW-YOLO achieves a 4.5% increase in mAP@0.5 and a 6.1% increase in recall compared to YOLOv8n. The generalization experiments demonstrate that the mAP@0.5 fluctuation of ASW-YOLO remains below 2% under strong highlight and striped shadow. Moreover, the model maintains over 85% mAP@0.5 under motion blur. ASW-YOLO balances precision and lightweight, making it suitable for real-time quality monitoring in industry.