基于文本图像互学习的换衣行人重识别方法  

Clothes-Changing Person Re-identification Method Based on Text-Image Mutual Learning

作　　者：葛斌[1] 卢洋 夏晨星 官骏鸣[2] GE Bin;LU Yang;XIA Chenxing;GUAN Junming(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001;School of Information Engineering,Huangshan University,Huangshan 245041)

机构地区：[1]安徽理工大学计算机科学与工程学院,淮南232001 [2]黄山学院信息工程学院,黄山245041

出　　处：《模式识别与人工智能》2024年第11期960-973,共14页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.62102003);安徽省自然科学基金项目(No.2108085QF258);安徽省博士后科学基金项目(No.2022B623);安徽理工大学研究生创新基金项目(No.2023CX2125)资助。

摘　　要：针对行人重识别在换衣场景下的小数据集样本中识别精度较低的问题,结合大模型CLIP(Contrastive Language-Image Pre-training)生成伪文本的功能,提出基于文本图像互学习的换衣行人重识别方法.在训练第一阶段,设计伪文本生成器,交换同批次中的样本像素,生成多样性文本,增强文本差异性,并通过语义对齐损失约束文本特征的一致性.在第二阶段,设计局部全局融合网络,融合局部特征和全局特征,在第一阶段文本信息的指导下,增强视觉特征的判别性.在PRCC、Celeb-ReID、Celeb-Light、VC-Clothes数据集上的实验表明,文中方法可提升在小数据集样本中的性能.To address the issue of low recognition accuracy in pedestrian re-identification(Re-ID)tasks involving clothing changes,a method for clothes-changing person re-identification based on text-image mutual learning(TIML)is proposed.It leverages the ability of contrastive language-image pre-training to generate pseudo-texts.In the first training phase,a pseudo-text generator is designed to enhance text diversity by swapping pixel information among samples within the same batch,thereby augmenting text variability.Additionally,a semantic alignment loss L SA is introduced to ensure the consistency in text feature representation.In the second phase of training,a global and local fusion network is devised to bolster the discriminative power of visual features by fusing local and global features,guided by the textual information obtained in the first phase.Experiments on PRCC,Celeb-ReID,Celeb-Light and VC-Clothes datasets demonstrate that the proposed model significantly improves recognition accuracy in scenarios with small dataset samples.

关 键 词：换衣行人重识别 CLIP 模态交互 语义对齐 提示工程 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]