Machine vision-based angle-arrayed imaging and two-stage deep learning for gear defect detection

Industrial gears are highly susceptible to surface defects under high-load and high-speed operating conditions, which can lead to reduced service life or even complete machine failure. However, complex operating environments pose significant challenges to achieving high-precision online defect detection. This paper proposes an online detection method integrating synchronous in situ tooth surface imaging with a two-stage deep segmentation strategy. The system achieves complete coverage of tooth surface images through motion-aligned imaging mechanisms and employs a two-stage cascaded architecture to enhance detection performance: the first stage rapidly segments the active tooth surface area, while the second stage utilizes an improved U-shaped dual-resolution network (UDDRNet) for precise identification of minute defects. Experimental results demonstrate that this method achieves 87.42% mIoU, 91.89% Recall, and 92.15% F1 scores while maintaining real-time performance, significantly outperforming single-stage methods and existing semantic segmentation models. These findings not only validate the high accuracy and practicality of the proposed method in complex industrial scenarios but also provide a scalable technical pathway for intelligent quality monitoring of critical industrial components.