Unlocking the Potential of Auxiliary Captions via Dual-Branch Multi-Scale Network for Composed Image Retrieval

Composed image retrieval (CIR) aims to retrieve target images by combining a reference image with a modification text. Traditional CIR methods often struggle with feature-level multimodal fusion, leading to deviations from the original embedding space. To address this, we propose a Dual-Branch Multi-Scale Network (DMN) that integrates a combining branch and a complete text branch. To enhance the use of captions generated by advanced image captioning models for CIR, the DMN leverages an attribute-driven disentanglement layer to separate features into distinct latent factors and employs a dual-path multimodal fusion module for effective feature integration. Additionally, a multi-scale matching module incorporating both global and local matching strategies is introduced to enhance fine-grained feature discrimination. Experimental results on the FashionIQ, Shoes, and CIRR datasets demonstrate that our DMN model consistently outperforms state-of-the-art methods, achieving improvements of up to 1.43% in mean recall metrics.