GSDD-net: an efficient feature extraction method for gear-surface defects in complex industrial environments

Abstract

Gear surface defects vary in scale and often exhibit subtle local features that are obscured by complex metallic backgrounds. However, achieving high detection accuracy and robustness with computationally lightweight models remains challenging in complex industrial inspection settings. We propose GSDD-Net, a lightweight framework based on the YOLO architecture for detecting gear surface defects. The proposed framework incorporates four task-oriented components: (1) ADown combines average and max pooling for downsampling to reduce model complexity while preserving the edge and texture details of subtle defects. (2) CSP-EDLAN is integrated into the backbone and neck networks. It uses cross-stage partial connections, multi-gradient path aggregation, and DualConv units to improve gradient propagation, cross-channel interactions, and multiscale feature fusion. (3) C2TSSA uses token statistical self-attention to capture global contextual dependencies with linear computational complexity, improving the representation of subtle defects against complex backgrounds. (4) LQEHead uses localization quality estimation to extract stable localization features from predicted bounding-box distributions and classification confidence scores. GSDD-Net achieves an mAP50 of 73.0% and an mAP50–95 of 38.0% on the real-world GID dataset. On a public gear defect dataset, it achieves 97.8% and 69.6%, respectively. Compared with the baseline model, GSDD-Net increases mAP50 and mAP50–95 on the GID dataset by 3.1 and 2.1 percentage points, respectively. GSDD-Net reduces the parameter count from 2.6 million to 1.9 million and the computational cost from 6.3 to 5.9 GFLOPs. It also achieves an inference speed of 105.7 FPS. These results indicate that GSDD-Net balances detection accuracy and computational efficiency under the evaluated conditions, supporting its potential for industrial gear surface defect detection.