Person Re-Identification Approach Utilizing Multi-Scale Information Fusion

To address the issue of feature loss in existing person re-identification methods, This paper presents a person Re-Identification network leveraging multi-scale feature fusion. Specifically, a multi-level scale information fusion module is proposed, utilizing the hierarchical structure of deep convolutional networks for fusion learning. This method allows the model to capture information at various scales from different network layer outputs, thereby enhancing the richness and representational power of the features. Additionally, an attention module is employed to refine the fused features, capturing both the essential and detailed features of pedestrians, thereby improving the model's capability to distinguish between pedestrian identities. This attention module can adaptively allocate the importance of features, increasing the network's focus on significant features, and thus improving recognition accuracy. Finally, the experimental results indicate that our method significantly improves the recognition performance of the model.achieving mAP and Rank-1 scores of 88.7% and 95.6% on the Market-1501 dataset, and 77.1% and 87.2% on the DukeMCMT-reID dataset, respectively.