Multi-Stage Edge-Guided Stereo Feature Interaction Network for Stereoscopic Image Super-Resolution

Stereo image super-resolution (SR) aims to simultaneously increase the resolution of stereo image pairs, which benefits many downstream three-dimensional (3D) multimedia broadcasting and stereo vision-related tasks, such as 3D television broadcasting and stereo matching. A key insight in convolutional neural networks-based stereo image SR is to enforce stereo feature interactions between the two stereo views to explore complementary cross-view features that can facilitate the SR in both views. To fully exploit the cross-view stereo features, in this paper we propose a new multi-stage network, cascaded by several stereo feature interactions, progressively improving the SR quality from coarse to fine. In particular, an edge-guided stereo attention mechanism is proposed to be embedded into each stereo feature interaction to better capture consistent structure details of the cross-views. Followed by stereo feature fusion and reconstruction modules, we finally put together a multi-stage edge-guided stereo feature interaction network (MESFINet) for stereo image SR. Comprehensive experiments on KITTI2012, KITTI2015, Middlebury, and Flickr1024 benchmark datasets show that the proposed MESFINet achieves superior performance against the state-of-the-art stereo image SR methods and can be used to improve the accuracy of stereo matching.