Cross-modality person re-identification based on multi-granularity fusion and cross-scale perception

A cross-modality person re-identification network based on multi-granularity fusion and cross-scale perception was proposed, which could effectively extract person image features and reduce the modality discrepancies between images. Firstly, a multi-scale feature fusion attention mechanism was proposed, and a multi-granularity non-local fusion framework was designed to effectively integrate image features from different modalities and scales. Secondly, a cross-scale feature information perception strategy was proposed, which could effectively reduce the influence of irrelevant noise caused by the change of perspective and person background on person discrimination. Finally, in view of the lack of person image feature information, a parallel dilated convolution residual module was designed to obtain more abundant person feature information. The proposed method was compared with current state-of-the-art cross-modal person re-identification algorithms on two standard public datasets. Experimental results show that the Rank-1 and mAP of the proposed method reach 75.9% and 73.3%, respectively, in the all search mode of the SYSU-MM01 dataset, and 93.7% and 89.3% in the VIS to IR retrieval mode of the RegDB dataset, respectively, which is better than the compared methods, which fully confirms the effectiveness of the proposed method.