Abstract
High-speed mass-production environments require reliable surface inspection because metal surface defects directly affect product reliability and process stability. Recent automatic surface inspection systems have increasingly adopted RGB image-based deep learning detectors to identify defects such as scratches, dents, and contamination. However, in practical inspection settings, the position and pose of small metal components can vary from frame to frame, and the reflective nature of metallic surfaces can make background texture visually similar to defect patterns. This study presents a cross-modal Region of Interest (ROI) generation framework that combines Time-of-Flight (ToF) intensity images with RGB images to define a stable component-level ROI before defect detection. The proposed method exploits signal attenuation and dropout patterns in ToF intensity images, caused by specular reflection on metallic surfaces, as structural cues for object localization. Foreground support regions extracted from the ToF intensity image are separated into object candidates using connected-component analysis and then projected onto the RGB coordinate system through a pre-calibrated homography. An object-adaptive ROI is generated from the projected region and provided to a YOLO-based defect detector. Experimental results show that the proposed method improves mAP@50 from 0.585 to 0.796 compared with the baseline YOLO26 model, demonstrating that ToF-guided object-adaptive ROI generation effectively reduces background interference and improves localization stability in metal surface defect detection.