Person Re-Identification Based on Local Salient Feature Embedding Network

Person re-identification is an important branch in the field of computer vision. In recent years, convolutional neural networks-based models, especially for the residual network, have made significant progress with the excellent ability to extract deep feature representation. However, these methods are not good at focusing on the local salient features related strongly to the target class effectively, which may limit the identification accuracy. To handle this issue, this paper proposes a local salient feature embedding network which contains a residual feature extractor, a local strong response pooling layer, a global average pooling layer, and a feature embedding module. Based on the deep residual features of pedestrians, the local salient features, which are crucial and special for different people, are obtained by the local strong response pooling. To increase the gap between inter-class variance and intra-class variance effectively, the feature embedding module integrates the local salient features and global semantic features and generate the low-dimension robust feature representation. Experimental results on the Market1501 and DukeMTMC-reID data sets demonstrate the effectiveness of the proposed model.