A Study on the Multi-Product Automatic Deburring Process Using SIFT, U2-Net-Based Vision Recognition, and Proxy-Controlled Path Optimization

This paper presents a vision-guided deburring automation system optimized for multi-variety production environments. The system randomly identifies input components using a combination of SIFT, template matching, and U²-Net algorithms. U²-Net, along with edge detection, was employed to detect burr regions precisely. The identified burr regions were processed through an optimized robot trajectory and refined using jerk minimization and collision avoidance techniques. To overcome the trajectory resolution limitations of the legacy Yaskawa robot controller, a proxy server architecture was introduced that enabled high-resolution path execution beyond native constraints. This integrated approach improves recognition accuracy, deburring efficiency, and automation flexibility, demonstrating the practical feasibility of labor-intensive die casting post-processing tasks in smart manufacturing.