Automated visual inspection (AVI) systems increasingly rely on deep learning-based object detection to deliver fast and consistent surface defect identification. However, the scarcity of defect image data and the shift in visual feature distributions caused by product variations remain significant challenges that hinder the practical deployment of deep learning systems. To address these challenges, we propose a synthetic defect generation pipeline based on an inpainting diffusion model. We introduce a Context-Aware Mask Generation (CAMG) module to ensure conforming mask generation required for realistic defect image synthesis. The proposed CAMG module employs a Self-DIstillation with NO labels (DINO) vision foundation model to automatically learn defect placement rules on product images. By synthesizing defects onto the non-defective images of new products via an inpainting approach, our method effectively bridges the distributional gap caused by product variations, enhancing robustness in New Product Introduction (NPI) scenarios. We validate the effectiveness of our method on footwear and printed circuit board (PCB) datasets. The experimental results demonstrate that incorporating the synthetic defect images consistently improves the mean average precision (mAP), precision, and recall performances of various defect detection models, with the most notable gains observed in the challenging NPI scenario.
Cho et al. (Fri,) studied this question.