基于多粒度区域相关深度特征学习的行人重识别  

Person Re-identification by Region Correlated Deep Feature Learning with Multiple Granularities

作　　者：董虎胜 钟珊[3] 杨元峰[1,2] 孙逊[1,2] 龚声蓉[3] DONG Hu-sheng;ZHONG Shan;YANG Yuan-feng;SUN Xun;GONG Sheng-rong(Jiangsu Province Support Software Engineering R&D Center for Modern Information Technology Application in Enterprise,Suzhou,Jiangsu 215104,China;School of Computer Engineering,Suzhou Vocational University,Suzhou,Jiangsu 215104,China;School of Computer Science and Engineering,Changshu Institute of Technology,Changshu,Jiangsu 215500,China)

机构地区：[1]江苏省现代企业信息化应用支撑软件工程技术研发中心,江苏苏州215104 [2]苏州市职业大学计算机工程学院,江苏苏州215104 [3]常熟理工学院计算机科学与工程学院,江苏常熟215500

出　　处：《计算机科学》2021年第12期269-277,共9页Computer Science

基　　金：国家自然科学基金(61702055,61972059,61773272);江苏省自然科学基金(BK20191474,BK20191475,BK20161268);苏州市职业大学校级科研项目(SVU2021YY03)。

摘　　要：在对行人重识别的研究中,联合使用从图像中提取的全身与局部特征已经成为当前的主流方法。但是许多基于深度学习的重识别模型在提取局部特征时忽略了它们在空间上的相互联系,当不同行人具有局部相似的外观时,这些局部特征的辨别能力会受到很大影响。针对该问题,提出了一种学习多粒度区域相关特征的行人重识别方法。该方法在对骨干网络提取的卷积特征张量作不同粒度的区域划分后,设计了区域相关子网络模块来学习融入空间结构关系的各局部区域特征。在区域相关子网络模块中,为了赋予局部特征与其他区域相关联的空间结构信息,综合利用了平均池化运算的空间保持能力与最大池化运算的性能优势。通过对当前特征和其他各区域的局部特征进行联合处理,使各局部特征间产生很强的空间相关性,提升了特征判别能力。在区域相关子网络模块的设计上,采用了与深度残差网络相同的短路连接结构,使得网络更易于优化。最后,由全身特征与使用区域相关子网络增强后的各局部区域特征联合实现行人重识别。Market-1501,CUHK03,DukeMTMC-reID 3个公开数据集上的实验结果表明,所提算法取得了优于当前主流算法的行人身份匹配准确率,具有非常优秀的重识别性能。Extracting both global and local features from pedestrian images has become the mainstream inperson re-identification.While among most of current deep learning based person re-identification models,the relations between adjacent body parts are seldom taken into consideration during extracting local features.This may decay the capability of distinguishing different persons when they share similar attributes of local regions.To address this problem,a novel method is proposed to learn region correlated deep features for person re-identification.In our model,the output feature map of backbone network is partitioned with multiple granularities first.And then the structure information preserved local features are learned via a new designed Region Correlated Network(RCNet)module.The RCNet makes full use of the structure maintenance of average pooling and the performance advantage of max pooling,endowing local features with rich structural information.By jointly processing current feature and local features from other regions,they are strongly related to each other due to the spatial correlation.As a result,the discrimination of them is significantly enhanced.For better optimization of the whole network,the shortcut connection in deep residual networks is also employed in the architecture of RCNet.Finally,the re-identification is conducted with both global features and the local features with structural information incorporated.Experimental results show that the proposed method achieves higher matching accuracies in comparison with existing approaches on the public Market-1501,CUHK03 and DukeMTMC-reID datasets,demonstrating favorable re-identification performance.

关 键 词：行人重识别 深度学习 特征表达 池化操作 区域相关网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]