A Novel Image Restoration Solution for Cross-Resolution Person Re-Identification

Cross-resolution person re-identification (CR-ReID) is a highly practical task that primarily addresses the image misalignment issue due to image resolution variations, which are caused by differences in the distances between cameras and camera performance variations. The existing cutting-edge approaches attempted to use super-resolution (SR) techniques to recover lost details in low-resolution (LR) images. However, the existing SR techniques focus on improving low-level semantic information metrics. Evaluation techniques for high-level semantic recognition tasks are not well suited for low-level image quality metrics. We propose a new framework called image feature restoration using a swin transformer (IFRSW), which uses the feature difference between LR and high-resolution (HR) images as the supervisory signal to constrain the SR module. We further improved the swin transformer by introducing a new multiresolution feature fusion strategy, enhancing its ability to extract features from multiresolution images. Additionally, we introduce a pooling technique called "softpooling" to preserve more information during the feature downsampling process. Our method exhibits a noteworthy 4.2% in rank 1 accuracy improvement on a challenging real LR dataset CAVIR, surpassing the current optimal approach. Our method achieves superior performance to the existing state-of-the-art (SOTA) methods.