Multi-scale Enhanced Fine-grained Feature-based Person Re-identification Algorithm

The key to solve the problem of Person Re-identification is to improve the extraction and application of Person effective features. Convolutional neural networks have powerful capabilities in this regard. This paper proposes a Person re-recognition algorithm based on multi-scale enhanced fine-grained features. Resnet50 is used as the backbone network to extract Person features at different scales, and the EFOM module is proposed to enable the extraction of fine-grained features by adding relevant global features while compensating for the shortcomings of its own attention mechanism to obtain enhancement and refinement. Finally, the MFFP module is used to obtain the fused features at different scales and then stitched into the BNNeck module. The fused feature vectors are supervised and trained using a variant triplet loss function with less overhead and a more flexible central loss function. Experimental results of the method on the DukeMTMC-ReID and Market-1501 datasets show that it achieves 86.7%% and 92.0% on the mAP evaluation metric; 91.1% and 94.8% on the Rank-1 evaluation metric. The experimental results show that the method makes full use of different scale feature information and key fine-grained features. It enhances the recognition degree of person features and improves the efficiency of person Re-ID.