Abstract:Text-based person search aims to localize target pedestrians from full-scene images based on natural language descriptions, which requires models to simultaneously possess global semantic understanding and local detail matching capabilities. Existing methods predominantly employ a single visual encoder, such as CLIP, to achieve cross-modal semantic alignment. While its self-attention mechanism exhibits powerful semantic alignment performance, the continuous multi-layer interactions tend to progressively smooth out local discriminative signals. To address this issue, a Dual-Branch Prototype Learning and Adaptive Feature Screening (DPLAS) method is proposed. Specifically, the global branch inherits the cross-modal semantic alignment capability of the Vision Transformer (ViT) to maintain robust global image-text matching. Concurrently, the local branch introduces ResNet-50, where the structured feature maps generated by its stage-wise downsampling provide precise spatial localization for part clustering. Meanwhile, the residual identity mapping continuously preserves the underlying detailed information unencoded by the global branch in the deep network, ultimately outputting a sequence of pixel features embedded with explicit spatial position information. These discrete features are then adaptively clustered into part-level semantic prototypes under the guidance of spatial soft constraints and diversity regularization. Furthermore, dynamic weight screening is applied to each prototype conditioned on the text query, effectively suppressing semantic collapse during the clustering process and filtering out background noise. Experimental results on three benchmark datasets demonstrate that the proposed method achieves a top Rank-1 accuracy of 75.63% on CUHK-PEDES, with improvements of 2.25%, 2.52%, and 1.90% over the baseline method, respectively, thereby validating its effectiveness in enhancing the accuracy of fine-grained cross-modal alignment.