Image-text semantic learning for unsupervised cross-resolution person re-identification

Cross-resolution person re-identification (CR-ReID) focuses on matching person images of the same identity across different resolutions. Most existing CR-ReID methods rely on manually annotated identity labels for training. Although some researchers have proposed unsupervised CR-ReID (UCR-ReID) methods, the feature fusion techniques they rely on still require a large number of parameters and significant computational resources, limiting the widespread application of UCR-ReID technology. To address the aforementioned issues, we propose an image-text semantic learning (ITSL) method, which incorporates text semantics to enhance recognition performance. During the testing phase, ITSL requires only a single encoder to obtain resolution-invariant features. Specifically, ITSL first learns text features based on a visual-language model, and then utilizes the dual semantic matching module to match inter-resolution positive clusters in both the image and text modalities. During the optimization process, ITSL not only incorporates image semantic contrastive loss to facilitate cross-resolution alignment but also integrates text semantic contrastive loss to leverage text semantics for promoting resolution-invariance learning. Additionally, we design random region downsampling in ITSL, which further enhances the model's robustness to resolution gaps through data augmentation. Experimental results on multiple cross-resolution datasets show that ITSL not only outperforms existing unsupervised methods while maintaining efficiency, but also approaches the performance of earlier supervised methods on certain datasets.