Local Feature Patch Matching Enhanced Re-ranking Model for Clothing Image Retrieval

Effective clothing image retrieval depends on robust global features and highly discriminative local features. We propose an effective re-ranking model to handle the issues of fine-grained difference discrimination and occlusion interference in clothing image retrieval. The global features are initially extracted by the Swin Transformer for a global-level retrieval, resulting in an initial set of candidate images. Then, a re-ranking network composed of multiple BiAttention layers is introduced to mine more discriminative local features from query-candidate image pairs. Each BiAttention module captures intra-image semantic information through self-attention, enhancing the representation capability for fine-grained differences. It also captures inter-image correlation by using bidirectional cross-attention to enhance the model's robustness to occlusion. In addition, the re-ranking network incorporates spatial relative attention to reinforce spatial position constraints. Positional importance weighting is further applied to matched feature blocks, so that the final similarity evaluation between image pairs appropriately concentrates on important feature regions. A large-scale clothing image dataset is built to validate the model's performance. Experimental results show that the proposed method is efficient in large-scale clothing image retrieval.