Domain adaptation via Transferable Swin Transformer for tire defect detection

Abstract

In data-driven tire defect detection, scarce defect samples and the poor transferability of models make the defect detection a time-consuming and expensive process. In recent years, domain adaptation methods have been applied to solve the cross-domain defect detection problems. However, previous works mainly assume that all regions in the image are equally transferable and align global features of the source and target domains with the same weight, which will fail for such a complex scenario. To this end, with the consideration of variability of transferability in different image regions, this paper presents Transferable Swin Transformer (TST), focusing the proposed method on region-based adversarial alignment and the semantic-based subdomain adaptation, simultaneously. Meanwhile, the subdomain adaptation benefits from the designed region-level weighting mechanism. The above strategies enable TST to learn both transferable and discriminative features in defect detection under domain shifts. Six transfer tasks constructed by different tire quality inspection X-ray image machines and a benchmark dataset are processed to demonstrate the effectiveness and generalization of the proposed method.