基于优化面部特征的换装行人重识别模型

打开文本图片集

中图分类号：TP391 文献标志码：A

Abstract：An optimized facial feature model （OFFM） based on ViT/B-16 was proposed to address the issue of clothing interference in cloth-changing person re-identification. The framework incorporated a spatial attention module and a facial-mask embedding module to refine facial features. Pedestrian features were fused and represented through locally optimized feature with global feature. Meanwhile， the features could fully utilize pedestrian facial information，resulting in improved recognition accuracy. In cloth-changing scenarios，enhancing facial perception and mitigating the interference caused by clothing variations substantially improved the identity discrimination ability of the model. The performance of OFMM was evaluated on public clothing change datasets，with ablation studies conducted to validate the model effectiveness. The experimental results show that the OFFM model improves mAP by 4.7% and Rank-1 by 4.1% on the PRCC dataset，and improves mAP by 5.2% and Rank-1 by 6.8% on the LTCC dataset，compared to the baseline models.

Keywords： person re-identification; deep learning；attention mechanism；computer vision

换装行人重识别（Cloth-changing person re-identification，CC-ReID）解决了换装场景下的身份匹配问题，解耦与服装无关的鲁棒性特征，在深度学习、机器人等领域发挥重要作用。（剩余7644字）