Defect-Aware Multi-View Adaptive Alignment for defect inspection in target-scarce domain adaptation

초록

Industrial visual inspection of 3D objects is challenged by the need to reliably detect subtle defects under distribution shifts caused by changes in process conditions and data acquisition settings. Multiple 2D images captured from different viewpoints provide a practical representation for this task, as they capture both geometric cues and fine-grained appearance details. However, defects are often visible in only a subset of views, and the number of defect-visible views varies across samples. Defect evidence is therefore distributed unevenly across views. In target-scarce domain adaptation, feature-level alignment provides a practical way to reduce cross-domain discrepancy while largely preserving the original inspection signal. However, when defect evidence is concentrated in only a few views, uniformly aligning multi-view representations can cause the adaptation process to emphasize defect-irrelevant variation, such as background differences, rather than defect-relevant information. To address this problem, we propose Defect-Aware Multi-View Adaptive Alignment (MAA), which assigns view-wise alignment weights based on defect evidence from each view. MAA derives semantic evidence from view-level classifier outputs and structural evidence from deviations from domain-specific normal references, and then calibrates the semantic evidence using predictive confidence. The resulting weights emphasize views with more salient defect cues. Experiments on a proprietary semiconductor dataset and the public Real-IAD dataset show that MAA consistently outperforms existing multi-view learning and domain adaptation methods. Compared with the strongest baseline, MAA improves average AUROC by 1.33 points on the semiconductor dataset and 1.24 points in the binary setting on Real-IAD.

키워드