Stereoscopic image super-resolution with interactive memory learning

Stereo image super-resolution aims to exploit the complementary information between image pairs and generate images with high resolution and rich details. However, existing methods explicitly calculate the similarity between image patches or pixels to build correspondence between different views. These hard-matching methods leave deep semantic information between image pairs unexplored. In this paper, a stereo image super-resolution method with interactive memory learning is designed to take advantage of the complementary information of different views in an implicit way. Specifically, we propose an interactive memory learning strategy to implicitly capture the semantic similarity between different views and design a feature dual-aggregation module for feature refinement. Extensive experiments on different datasets achieve state-of-the-art results, demonstrating that our method effectively boosts the quantitative and qualitative results of stereoscopic image pairs. Code can be found at: https://github.com/zhuxiangyuan1/IMLnet.