A study of person re-identification approach utilizing an enhanced ConvNeXt architecture

In this paper, we present a new person re-identification method based on an improved ConvNeXt network, called ConvNeXt-AP. This method effectively captures pedestrian features, reduces computing resources, and improves recognition accuracy. ConvNeXt is used as the main network to enhance the capture of local spatial features, with the final forward head module removed to retain more pedestrian features. Moreover, the segmentation strategy from the PCB_RPP network is incorporated at the end of the model to extract fine-grained information from the pedestrian image. To improve the effectiveness of feed-forward convolutional neural networks, the ConvNeXt model Block implements the non-parametric attention mechanism SimAM. This mechanism infers the three-dimensional attention weight of the feature map without adding parameters to the original network, resulting in significant performance improvement with minimal maintenance overhead. To achieve stability and faster convergence, the model achieves stability over time by using a warm-up strategy, and during training, a random erasing strategy is used to reduce the risk of overfitting and improve robustness to occluded pedestrians. Our method has been rigorously tested on multiple public datasets of person re-identification, and the results demonstrate superior performance compared to many state-of-the-art methods.