Research on fashion object detection based on cross-domain adaptive learning

Abstract

Accurate and scalable recognition of fashion image attributes is essential for the digital transformation of the apparel industry. However, fashion images are often highly stylized, significantly impacting the accuracy of detection and classification. In this paper, we present a multi-object cross-domain adversarial network with detection enhancement (MCAN-DE) model for fashion object detection. Our model uses multi-label predictions to identify object categories and applies conditional adversarial global feature alignment to balance features across domains while preserving their discriminability. Additionally, we introduce a prediction consistency regularization mechanism that uses multi-label predictions to maintain consistency in object category detection across domains. Experimental results show that, compared with the source-only model, the proposed method achieves an average accuracy improvement of 16.1 % on the stylized artwork image datasets and 11.2 % on the stylized fashion datasets, corresponding to a relative gain of over 20 %. Notably, the “trousers” and “skirt” categories reach precision rates of 71.3 and 66.4 %. Our work extends the scope of fashion object detection and offers promising avenues for the apparel industry’s intelligent transformation.