Focused on the issue that huge modal difference between cross-modal person re-identification images， pixel alignment and feature alignment are commonly utilized by most of the existing methods to realize image matching. In order to further improve the accuracy of matching two modal images， a multi-input dual-stream network model based on dynamic dual-attention mechanism was designed. Firstly， the neural network was able to learn sufficient feature information in a limited number of samples by adding images of the same person taken by different cameras in each training batch. Secondly， the gray-scale image obtained by homogeneous augmentation was used as an intermediate bridge to retain the structural information of the visible light images and eliminate the color information at the same time. The use of gray-scale images weakened the network’s dependence on color information， thereby strengthening the network model’s ability to mine structural information. Finally， a Weighted Six-Directional triple Ranking （WSDR） loss suitable for images three modalities was proposed， which made full use of cross-modal triple relationship under different angles of view， optimized relative distance between multiple modal features and improved the robustness to modal changes. Experimental results on SYSU-MM01 dataset show that the proposed model increases evaluation indexes Rank-1 and mean Average Precision （mAP） by 4.66 and 3.41 percentage points respectively compared to Dynamic Dual-attentive AGgregation （DDAG） learning model.

Whether the learned features have better invariance in the significant intra-class changes will determine the upper limit of performance of the Person Re-identification （ReID） model. Environmental light， image resolution change， motion blur and other factors may cause color deviation of pedestrian images， and these problems will cause overfitting of the model to color information of the data， thus limiting the performance of the model. By simulating the color information loss of the data samples and highlighting the structural information of the samples， the model was helped to learn more robust features. Specifically， during model training， the training batch was randomly selected according to the set probability， and then a rectangular area of the image or the entire image was randomly selected for each RGB image sample in the selected batch， and the pixels of the selected area was replaced with the pixels of the same rectangular area in the corresponding grayscale image， thus generating a training image with different grayscale areas. Experimental results demonstrate that compared with the benchmark model， the proposed method achieves a significant performance improvement of 3.3 percentage points at most on the evaluation index mean Average Precision （mAP）， and performs well on multiple datasets.