Abstract:

Surface defect detection is very important to ensure the quality of industrial products. Traditional machine learning cannot be well extended to a non-identically distributed dataset, making surface defect detection a data-limited task. Unsupervised domain adaptation (UDA) can solve this problem by transferring knowledge from a labeled source domain to an unlabeled target domain. Adversarial learning is one of the latest heuristic methods to deal with domain shift in UDA tasks. Although impressive results have been achieved, the adversarial model still suffers from the equilibrium challenge, which may lead to under-transfer or negative transfer. To this end, we utilize joint distribution adaptation to propose a novel UDA model, named significance-augmented-based category-level selective dual-adversarial (CDASA) network, to learn a generalized model. Specifically, to promote positive transfer, we use the selective dual-adversarial (DA) learning strategy to further minimize the feature distribution difference at equilibrium to achieve better domain confusion. Meanwhile, guided by defect recognition performance, the transferability of confusion domain is measured to enhance the distributions of potential domains. In addition, to avoid under-transfer, we consider the relationship between the target data and the decision boundary, and then the significance-augmented (SA) mechanism is proposed to encourage class-level alignment. Thus, the alignment features with domain-invariant and category discrimination can be captured simultaneously. Extensive experiments on collected real industrial datasets and publicly available steel surface defect datasets confirm the effectiveness of our approach.

This article is available online here and to CALCE Consortium Members for personal review.