Person Re-identification Lightweight Network Based on Progressive Attention Mechanism

This paper proposes a lightweight person re-identification network that incorporates a progressive attention mechanism The network aims to address the low accuracy issue in person re-identification caused by various factors such as different viewing angles, poses, illumination conditions, occlusions, and low image resolutions. Additionally, the network design takes into consideration the need for lightweight model deployment in practical scenarios. To improve the network’s performance using limited training data, data augmentation techniques are employed to expand the training dataset and enhance the robustness of the network model. The Resnet-50 architecture serves as the backbone network, and a feature shunt structure with depthwise separable convolutions is introduced to reduce computational parameters and accelerate person retrieval inference speed. Furthermore, the feature extraction and embedding processes are separated, and a progressive attention module is introduced. This module gradually segments the features into local blocks of different granularity, allowing for the learning of discriminative features at each granularity level. This progressive approach enhances the network’s ability to perceive foreground information from coarse to fine levels and improves feature matching capability. To supervise the model, a triplet loss function is utilized, specifically designed to address challenging samples. This loss function helps reduce intra-class variations while increasing inter-class separability. The efficacy of the proposed method in person re-identification is substantiated by conducting experimental evaluation on both the Market-1501 and DukeMTMC-ReID datasets. The experimental results demonstrate that the method achieves mAP indices of 88.1% and 79.1% on the respective datasets, providing strong evidence for its effectiveness in addressing the challenges of person re-identification.