Learning Multi-dimensional Parallax Prior for Stereo Image Super-Resolution

Recent years have witnessed great success in image super-resolution (SR). However, how to effectively exploit stereo information for the SR purpose is still challenging. This paper focuses on proposing a general solution to stereo image SR. We propose a novel module named Parallax Multi-Dimensional Attention (PMDA) that could not only be seamlessly integrated into most of existing SISR networks but also explore cross-view information from stereo images. Specifically, a pair of stereo images are fed into two identical SISR networks. The extracted middle features are transferred into PMDA to capture the inherent correlation within stereo image pairs. Finally, the internal-view and cross-view information is mixed by SISR network to generate the final output. We also introduce Self Multi-Dimensional Attention (SMDA) to effectively improve the feature representation capacity of a single image. Based on PMDA and SMDA, we design a stereo image SR model named Progressive Attention Stereo SR (PASR), which illustrates the flexibility of PMDA and performance-boosting guided by PMDA and SMDA. Extensive experiments show its superiority in the aspects of visual quality and quantitative comparison.