Semi-supervised adaptive network for commutator defect detection with limited labels

Abstract

Deep learning-based surface defect detection methods have obtained good performance. However, customizing architectures for specific tasks is a complex and laborious process. Neural architecture search (NAS) offers a promising data-driven adaptive design approach. Yet, deploying NAS in industrial applications presents challenges due to its reliance on supervised learning paradigm. Hence, we propose a mixed semi-supervised adaptive network for commutator surface defect detection, even with limited labeled samples. In the proposed framework, we employ a multi-branch network with complementary perturbation flows, leveraging consistency regularization, pseudo-labeling, and contrastive learning. First, a confidence-guided directional consistency regularization strategy aligns features in high-quality directions. Second, confidence-aware hybrid pseudo-labeling improves the pseudo-supervision quality. Finally, foreground/background contrast awareness encourages the model to more sensitively identify defect regions. The detection backbone is data-driven generated through a neural architecture search process, replacing manual design strategies. Experimental results show our method automatically generates optimal commutator detection networks using limited labels, outperforming existing state-of-the-art methods. Our work paves the way for adaptive defect detection networks with limited labels and can extend to surface defect detection in various production lines.