Subdomain adaptation network with category isolation strategy for tire defect detection

Data-driven tire defect detection has been greatly confronted with the scarcity of defective samples and the poor transferability of models. One of the best ways to alleviate these issues is domain adaptation. However, the tire defect features have large intra-class differences and small inter-class differences. Therefore, most of the existing domain adaptation methods that only concentrate on global distributions will fail for such a complex scenario. To overcome the weakness, this article first proposes the multi-representation-based subdomain adaptation network with category isolation strategy, and then describes the category isolation strategy and multi-representation-based alignment. With the consideration of fine-grained distributions of two subdomains, the designed category isolation strategy makes the features of different subdomains more discriminative. Besides, an Inception module is introduced to extract multi-representation of the complex tire X-ray images. Finally, extensive experiments conducted on the tire defect dataset and a benchmark dataset demonstrate the effectiveness of the proposed method.