Capturing Cross-Scale Disparity for Stereo Image Super-Resolution

Abstract

Stereo image super-resolution (SR) exploits the stereo feature information from cross-view image pairs for image resolution. This paper focuses on how to effectively exploit the disparity information between stereo viewpoints and proposes a cross-scale parallax-attention network (CSPAN) for stereo image SR. Specifically, a novel cross-scale parallax attention module (CPAM) is developed to explore cross-scale parallax prior. Moreover, instead of the widely used upsampling module based on the sub-pixel layer, we present a novel cascade dynamic upsampling module (CDUM), which not only dynamically generates the upsampling filters according to the input content, but also restores the high frequency details in a coarse-to-fine manner. Especially, nonlinear activation free blocks (NAFBlocks) are used as the feature extraction module in our network, which further improves the performance of our model. Extensive experiments on Mid- dlebury, KITTI 2012 and KITTI 2015 demonstrate that the proposed framework outperforms many competitive stereo SR methods in both PSNR and SSIM. Code is available at https://github.com/DoragonKuesuto/CSPAN.