MixStyle-Based Dual-Channel Feature Fusion for Person Re-Identification

The problem of Person Re-Identification is still a big challenge, as the complex network structure and unsatisfactory generalization performance of widely used deep neural networks make them unsuitable for application to real-world problems. In this paper, we propose a global feature-based person re-identification network with strong generalization. The extracted features part contains two channels of feature fusion: the feature extraction module and the feature generalization module. The feature generalization module is a new MixStyle module added to the feature extraction module, which can effectively mix the style information of images under different domains or even the same domain to form multiple potential domain features, thus improving the generalization performance of the model. In addition, this paper also makes some improvements to the loss function by adding a new constraint on the positive sample pair distance, which makes it possible to maximizes the reduction of intra-class distance in addition to pushing the distance between different classes during the training process. Experimental results on two datasets, Market1501 and DukeMTMC, demonstrate that the method proposed in this paper exhibits strong generalization performance for the person re-identification problem and outperforms current global feature-based person re-identification methods.