A Hybrid Feature Integration Network for Robust Person Re-Identification

In the task of person re-identification (ReID), cross-view and cross-scale feature representation is crucial for model accuracy. To address the limitations of traditional deep learning methods in feature representation and input robustness, this paper proposes a novel Hybrid Feature Integration Network (HFI-Net), which combines Convolutional Neural Networks and Vision Transformers. It introduces a Unified Integration Module (UIM) and an Enhanced Fusion Attention Module (EFAM). UIM implements a weighted mechanism for multi-view feature fusion, while EFAM, based on channel and spatial attention mechanisms, integrates convolution and feature mixing operations to enhance feature discrimination. By jointly optimizing multiple loss functions, HFI-Net efficiently integrates multi-view and multi-scale information. Experiments on mainstream datasets such as Market-1501 and DukeMTMC-ReID demonstrate significant improvements in Rank-1 and mAP, validating the model’s effectiveness and robustness.