Person Re-identification Based on Dual-channel Attention and Local Feature Fusion

Abstract:
Aiming at improving low accuracy of existing person re-identification algorithms in dealing with low image resolution, different illumination, diverse poses, and viewing angles, this paper proposes a person re-id algorithm based on dualchannel attention and local feature fusion. After ResNet50, it is followed by the spatial attention and the channel attention to extract the global features. Then the feature map segmentation is performed to extract the local features, and a fusion feature with strong judgment is obtained. In addition, this paper adds the central loss constraint item to the triplet hard loss. The triplet_center loss joint cross-entropy loss function further to improve adaptability of the model in complex scenarios. The proposed method is compared and tested on the CUHK03, DukeMTMC-reID and Market-1501 data sets, and the accuracy of Rank-1 reached 96.1%, 89.8% and 79.8%. This algorithm makes full use of the features of different depths and the key information of interest. The module has a solid discriminative ability. Also the accuracy of mean average precision of the proposed method is better than other advanced algorithms.