基于特征细化的多标签学习无监督行人重识别

Multi-label learning unsupervised person re-identification based on feature refinement

针对无监督行人重识别中行人特征表达不充分以及训练过程产生噪声标签等问题,提出了一种基于特征细化的多标签学习无监督行人重识别方法。首先,为提高网络对关键区域信息的利用能力,设计多尺度通道注意力模块(Multi-scale channel attention module,MCAM),嵌入ResNet50网络的不同层来构建特征细化网络,并利用该网络对输入图像通道维度上的关键信息进行强化和关注,以获得更丰富的特征信息;其次,为降低训练过程中产生的噪声标签对网络的负面影响,设计多标签学习模块(Multi-label learning module,MLM),通过该模块进行正标签预测以生成可靠的伪标签;最后,利用多标签分类损失和对比损失进行无监督学习。在数据集Market-1501和DukeMTMC-reID上进行实验,结果表明该方法在这两个数据集上的平均精度均值分别达到82.8%和70.9%,首位命中率分别达到92.9%和83.9%。该方法使用注意力机制强化图像的特征信息,并通过正标签预测减少噪声标签,有效提升了无监督行人重识别的准确率,为无监督行人重识别领域提供了更鲁棒的方法。

Aiming at the issue of inadequate expression of person features and noise labels generated in the training process in unsupervised person re-identification,we proposed a multi-label learning unsupervised person re-identification method based on feature refinement.Firstly,to improve the network's ability to use key area information,a multi-scale channel attention module(MCAM)was designed.We embedded it into different layers of ResNet50 to construct a feature refinement network.This network was used to srengthen and focus the information on the channel dimension of the input image to obtain richer feature descriptions.Secondly,to reduce the detrimental effects of noise labels produced during network training,we designed a multi-label learning module(MLM).Positive label prediction was performed through this module to generate reliable pseudo-labels.Finally,unsupervised learning was carried out by using multi-label classification loss combined with contrast loss.We conducted experiments on Market-1501 and DukeMTMC-reID datasets.The results show that the Rank-1 hit rate is 92.9%and 83.9%,while the mean average precision reaches 82.8%and 70.9%,respectively.This method uses the attention mechanism to enhance the feature information of the image and reduces the noise label by positive label prediction.It effectively improves the accuracy of unsupervised person re-identification and provides a more robust method for unsupervised person re-identification fields.