损坏图像下基于风格归一化与全局注意力的行人重识别

Pedestrian re-identification based on style normalization and global attention in corrupted images

针对当前网络难以应对各种损坏类型的行人图像与易丢失跨维信息的问题,提出了一种损坏图像下基于风格归一化与全局注意力的行人重识别(pedestrain re-identification,ReID)方法。该方法通过平滑极大单元的风格归一化与恢复(smooth maximum unit-style normalization and restitution,SM-SNR)模块中的实例规范化(instance normalization,IN)过滤掉域中的风格变化,同时平滑极大单元(smooth maximum unit,SMU)能使该模块更充分地从删除的信息中提取行人相关特征并将其恢复至网络中,缓解损坏图像带来的风格差异。此外,全局注意力机制(global attention mechanism,GAM)通过关注通道与空间之间的相互作用,以捕获3个维度上的显著行人特征,减少跨维信息的丢失,最终使本模型在面对行人损坏图像时的识别能力得到有效提高,且保留了在干净数据集上的竞争力。实验结果表明,本算法在损坏测试集上的各项指标与目前主流算法对比具有显著的优越性。其中,本模型与2021年的CIL模型使用CUHK03数据集比较的结果为:在Corrupted Eval上,R-1、mAP和mINP分别提高了15.18%、15.75%与11.65%;在Clean Eval上,R-1与mINP仅降低了0.24%、0.75%,mAP提升了0.25%。

Aiming at the problem that the current network is difficult to deal with various corrupted pedestrian images and easily loses cross-dimensional information,a pedestrian re-identification(ReID)method based on style normalization and global attention is proposed for corrupted images.The method filters out style changes in the domain by smooth maximum unit-style normalization and restitution(SM-SNR)module in the instance normalization(IN),and at the same time smooth maximum unit(SMU)enables the module to more fully extract pedestrian-related features from the deleted information and restore them to the network,so as to alleviate the style difference caused by corrupted images.In addition,the global attention mechanism(GAM)captures the salient pedestrian features in three dimensions by focusing on the interaction between the channel and the space,reducing the loss of cross-dimensional information.Finally,the recognition ability of the model in recognizing pedestrian corrupted images is effectively improved,and the competitiveness on clean datasets is retained.The experimental results show that the indicators of the algorithm on the corrupted test set has significant advantages compared with the current mainstream algorithms.Among these algorithms,the result of comparison with the 2021 CIL model using the CUHK03 dataset is that:On Corrupted Eval,R-1,mAP and mINP increase by 15.18%,15.75%and 11.65%respectively;on Clean Eval,R-1 and mINP only decrease by 0.24%,0.75%,and mAP increased by 0.25%.