Local-enhanced representation for text-based person search

Text-based person search is a critical task in intelligent security, designed to locate a person of interest by text descriptions. The primary challenge in this task is to effectively bridge the significant gap between the text and image domains while simultaneously extracting the discriminative features that are crucial for the accurate identification of individuals. Existing methods have made some effective attempts by conducting cross-modal matching at the fine-grained representation level. However, these approaches frequently overlook two crucial factors: (i) the presence of noise in the local features during information fusion, and (ii) the lack of intra-modal matching when measuring feature similarity. To address the above issues, we propose a novel local-enhanced representation framework in this paper. Specifically, to restrain noises in local features, we design a Relation-based cross-modal local-enhanced fusion module, which can filter out weak related information by relation assessment. In addition, we explore an intra-cross modal projection strategy to overcome the limitations of existing cross-modal projection methods. This strategy jointly applies the intra-modal and cross-modal matching constrains in feature distribution. Finally, experiments on three mainstream datasets verify the performance superiority of our proposed method compared to existing state-of-the-art methods.