YOLO-RCHA: a gear defect detection method based on robust context-aware hybrid attention mechanism

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Gear surface defect detection plays a crucial role in gear inspection, quality control and maintenance. In the process of defect detection, the complex and variable shooting situations and the multiscale, small-sized, complex texture, low-contrast defects lead to inefficient and inaccurate defect detection. To address the issues, a high-efficiency gear surface defect detection approach is presented based on the YOLOv5-v7.0 network, named YOLO with Robust Context-aware Hybrid Attention (YOLO-RCHA). First, a new C3 with Global Grouped Coordinate Attention Pyramid (C3GGCAP) module is proposed to enhance the local information extraction capability. Then, to eliminate the interference of noise and outliers, the C3 with Median-Enhanced Channel Spatial Attention (C3MECSA) module is introduced in the neck, which enables efficiently to construct a wide receptive field, and further enhances the understanding of global features. Afterwards, the Global Channel Shuffle Spatial Attention (GCSSA) module is embedded before head to amplify the cross-dimensional feature interaction and improve the model’s capacity to identify hidden or low-contrast defects. The experimental results illustrate that the improved YOLO-RCHA network presents a better overall performance with mAP@0.5 of 97.0% and 75.7% on NEU-GSD and NEU-DET datasets, which increase by 0.8% and 1.6% than that of YOLOv5-v7.0 network, respectively.