Gear Defect Detection Method Based on CS-Attention and Faster-RCNN

Gear defect detection is critical in the manufacturing of mechanical components but is often hindered by complex backgrounds, diverse defect types, and difficulty detecting small defects. To address these challenges, we propose a gear defect detection method based on ViT and Faster-RCNN. This method incorporates a CS-attention structure that combines HLattention, which integrates global and contextual weights, with sparse attention to reduce computation costs and enhance multi-feature fusion. Additionally, a feature pyramid network (FPN) is used to efficiently fuse multi-level features. The Faster-RCNN neck network and detection head process these fused features. Experimental results on the Guizhou University gear defect dataset show that the method achieves 93% average precision and 30 FPS detection speed, offering high accuracy with lower computational cost, making it suitable for industrial gear surface defect detection.