Multi-View Attention Guided Feature Learning for Unsupervised Surface Defect Detection

Unsupervised surface defect detection is gaining attention in quality assurance of industrial manufacturing. However, existing methods typically rely on single-view data, which will lead to imaging blind spots and poor discrimination of subtle defects due to the complex structures of industrial components. Therefore, this article proposes a novel multiview attention guided feature learning (MVAG-FL) approach, which offers two primary advantages. First, the MVAG-FL leverages intra-view and cross-view attention to adaptively integrate in-formation and effectively learns subtle defect features from mul-tiple perspectives. Second, a multilevel dual perturbation module, combining layer-wise jitter and bottleneck perturbation, is proposed to improve the robustness and performance of the MVAG-FL in a multiclass setting. Considering the requirement for global feature consistency of different perspectives and focusing on local defect regions, a global-local adaptive cosine loss is designed to align multiview features and amplify the difference between subtle and normal features. Experiments were conducted on the Real-IAD dataset with the multiclass setting, and the results show MVAG-FL outperforms the state-of-the-art multiview detection methods, with 1.2% image-level and 4.5% pixel-level AUROC improvements.