用于车辆重识别的视角感知局部注意力网络

View-Aware Part Attention Network for Vehicle Re-Identification

车辆重识别的目的是从大型车辆数据库中找到与查询车辆相同特征的所有车辆图片。目前,由于同一车辆在不同视角下外观差异大或颜色、车型相同的不同车辆在特定视角下外观差异小,导致车辆重识别的准确度和鲁棒性均有待提高。提出一个视角感知局部注意力网络,采用弱监督注意力学习方式代替人工手动的车辆局部部件标注,自适应学习每个视角内所有显著性局部特征。通过局部注意力裁剪操作裁剪并放大该视角领域内部件细节信息,并基于局部注意力擦除操作擦除一些局部区域,以鼓励模型发掘该视角领域内其他更多的显著性局部线索。构建一种共同视角的注意力增强模块,以强化共同视角特征学习,并根据视角的相似度给每个视角分配相应的权重,使同一视角特征学习得到增强,不同视角特征学习受到抑制。实验结果表明,所提网络在VeRi-776数据集下的mAP为81.2%,在VehicleID数据集下的CMC@1、CMC@5分别为85.7%、98.0%,相较于PRN、PVEN、SAVER等重识别网络具有更高的识别精度和更强的泛化能力。

Vehicle re-identification aims to retrieve all same-identity images from querying vehicle images from a large image database.Currently,the appearance difference of same vehicles under different perspectives is large,whereas the appearance difference of different vehicles under specific perspectives is small probably due to the having same color and model,which leads to a need for the improvement of the accuracy and robustness of vehicle image recognition.A View-Aware Part-Attention Network(VPAN)is proposed,and a weakly-supervised attention-learning method is used to replace manual vehicle local component labeling to adaptively learn all significant local features in each perspective.The detail information of the internal parts in a perspective field is clipped and enlarged by a local attention-clipping operation,and some local areas are erased based on a local attention-erasing operation to encourage the model to discover more significant local clues in the perspective field.A common perspective attention enhancement module is constructed to strengthen common perspective feature learning.Each perspective is assigned a corresponding weight according to the similarity of perspectives,so that same perspective feature-learning is enhanced and different perspective feature-learning is suppressed.The experimental results show that a map of the proposed network with the viri-776 dataset is 81.2%,and that with the vehicleid datasets CMC@1 and CMC@5 are 85.7%and 98.0%respectively.Compared with PRN,PVEN,SARATR and other re-identification networks,the proposed network has higher recognition accuracy and generalization ability.