Clothing Change Person Re-Identification Method Based on Facial Feature Assistance

To address the inadequacies of conventional person re-identification (ReID) approaches in handling clothing variation scenarios and the challenges arising from partial facial occlusion or absence, this paper presents a novel facial feature-augmented ReID framework. Confronted with the constrained feature extraction capability of existing Cross-Attention Multi-Objective Comparison (CAMC) networks in capturing comprehensive pedestrian attributes, our methodology introduces a dedicated facial recognition branch. Specifically, we implement a multi-stage facial detection and alignment process through the MTCNN module, subsequently establishing a hierarchical fusion framework that synergistically combines facial biometric patterns with chromatic (RGB) and anthropometric (body shape) features derived from the CAMC backbone. This integration facilitates the generation of robust cross-modality representations that significantly improve discriminative capability. Moreover, we devise a cross-modal attention mechanism for multi-scale feature aggregation, enabling comprehensive interaction between localized facial semantics and global pedestrian characteristics. Extensive experiments reveal substantial performance gains over the baseline CAMC model: The proposed method achieves absolute improvements of 1.0% mAP, 4.8% Rank-1, and 7.9% Rank-5 accuracy on the LTCC dataset, while demonstrating more pronounced enhancements of 28.9% mAP, 22.6% Rank-1, and 14.9% Rank-5 accuracy on the Celeb-ReID benchmark. These statistically significant advancements across multiple evaluation metrics substantiate the technical superiority and practical efficacy of our approach in addressing clothing-agnostic person recognition challenges.