Leveraging more information for blind stereo super-resolution via large-window cross-attention

Abstract

Stereo image super-resolution aims to reconstruct high-resolution images by effectively utilizing cross-view complementary information from stereo image pairs. A prevalent method in Stereo image super-resolution is the stereo cross-attention module, which allows the model to focus on and integrate relevant features from both the left and right views. Despite its advantages, our analysis using a diagnostic tool called local attribution map (LAM) reveals that current methods exhibit limitations in effectively leveraging this complementary information. To address this issue, we propose the Double Stereo Cross-Attention Module (DSCAM), which utilizes an Overlapping Stereo Cross-Attention (OSCA) mechanism that enhances the integration of cross-view complementary information by using overlapping windows, followed by an additional multiplication step to refine and emphasize the combined features. Additionally, we develop a stereo image degradation model that ensures the consistency of degradation between stereo pairs, accurately simulating the real-world degradation process of stereo images. Extensive experiments have demonstrated that our method achieves visually pleasing results, making it the first to address the problem of stereo image super-resolution in real-world scenarios. The source code is available at https://github.com/nathan66666/LCASSR.git.