行人再识别中的多尺度特征融合网络

Multi-scale Feature Fusion Network for Person Re-identication

针对行人再识别中待识别对象和目标对象的体态、衣服的颜色等外貌特征非常相似时,模型难以正确识别行人身份这一难点问题,提出了一个基于残差网络ResNet50改进的多尺度特征融合网络.通过利用最后一层特征协同多个中间层特征,采用顶层到下层递进式加和的特征层融合机制来提取行人图像特征,确保模型在总体特征表述基础上,提高对微小细节信息的表征能力.在3个主流的行人再识别公共数据集Market-1501、CUHK03(D)和DukeMTMC-reID上进行了实验,与2018年同类型的行人再识别网络DaRe相比,提出的方法比Market-1501数据集的Rank-1指标提升了2.82%,mAP指标提升了4.32%;比DukeMTMC-reID数据集的Rank-1指标提升了5.45%,mAP指标提升了6.4%.实验结果证明了所提出方法的有效性.

In view of the similarities between the appearance of the object to be identified and the color of the object in the person re-identification,the problem with the model is that it is difficult to correctly identify the pedestrian's identity.To solve this problem,a multi-scale feature fusion network based on the residual network ResNet50 was proposed,in which the last layer feature was used to cooperate with multiple intermediate layer features.The top-level to lower-level progressive addition feature layer fusion mechanism was adopted,and the pattern of the pedestrian image ensured that the ability to represent small details in this mode was improved based on the overall feature representation.Experiments were carried out on three mainstream person re-identification public datasets Market-1501,CUHK03(D)and DukeMTMC-reID respectively.Compared with the same type of person re-identification network DaRe in 2018,the proposed method in Rank-1 indicator on the 1501 dataset increased by 2.82%,and the mAP indicator increased by 4.32%;the Rank-1 indicator of the DukeMTMC-reID dataset increased by 5.45%,and the mAP indicator increased by 6.4%.The experimental results show that the proposed method is effective.