结合注意机制和局部擦除的行人重识别方法

Person Re-identification Method Based on Attention Mechanism and Local Erasure

针对行人重识别中行人姿态变化和遮挡问题,提出了一种结合注意机制和局部擦除的行人重识别方法。首先,构建由ResNet50为全局分支和注意擦除为局部分支组成的双分支网络。全局分支用来提取全局特征表示,在训练过程中可以监督注意擦除分支的训练。注意擦除局部分支由注意模块和擦除模块组成,该分支将输入特征映射的同一区域随机地分批擦除,以增强局部区域的注意特征学习;其次,在训练阶段采用标签平滑损失函数和三元组损失函数对模型进行联合训练。标签平滑损失函数用于防止分类任务过度拟合,三元组损失函数用于解决类间相似、类内差异的分类问题;最后,在Market-1501,DukeMTMC-reID两个数据集上、和现有的八种方法进行对比测试实验,rank-1/mAP分别达到94.3%/85.9%,87.2%/75.3%,优于其他现有方法。

Aiming at the problem of pedestrian pose change and occlusion in person re-identification,a person re-identification method combining attention mechanism and local erasure is proposed.Firstly,a dual branch network is constructed,which consists of resnet50 as global branch and erasure as local branch.The global branch is used to extract the global feature representation,and it can supervise the training of the attention erasing branch during the training process.The local branch of attention local erasure consists of an attention module and an erasure module,which erases the same region of the input feature map randomly in batches to enhance the local region attention feature learning;secondly,the label smoothing loss function and the triple loss function are used to jointly train the model in the training stage.The label smoothing loss function is used to prevent the over fitting of classification tasks,and the triplet loss function is used to solve the classification problems of similarity between classes and differences within classes.Finally,a comparative test is conducted on market-1501 and DukeMTMC-reID data sets,and the Rank-1/mAP is respectively 94.3%/85.9% and 87.2%/75.3%,which is better than other existing methods.