Enhancing Person Re-Identification with TF-SS: A Transformer Model based on Shifted Windows Division and Skipping Sub-Layer Techniques

Person re-identification (Re-ID) involves recognizing a specific individual from images captured across different devices, which has splendid application prospects. Due to factors such as motion changes and target occlusion, the accuracy of person re-identification algorithms in complex scenes will be reduced, and existing algorithms fail to solve such problems properly. On this basis, this paper proposes a person re-identification model named Transformer Shifted Window with Skipping Sub-Layer (TF-SS) based on the shifted window division and skipping sub-layer methods. Firstly, a relative position encoding structure is introduced into the Transformer to improve the model's ability to capture spatio-temporal features. Secondly, the standard multi-head self-attention module of the Transformer is replaced with a multi-head self-attention module based on shifted windows, which can perform better dense predictions on hierarchical feature maps with linear computational complexity. Then, an adversarial training algorithm is introduced into the semantic embedding layer to enhance the robustness of the model. Finally, the skipping sublayer method is introduced. By randomly omitting sub-layers to introduce perturbations into the training, a greater constraint effect is exerted on the sub-layers, further enhancing the accuracy of the model. In this paper, the effectiveness of TF-SS is verified through extensive ablation experiments and comparative analyses, and it outperforms the current most advanced person re-identification methods in terms of mean Average Precision (mAP) and Rank@5 and Rank@10 accuracies.