Stereo Cross Global Learnable Attention Module for Stereo Image Super-Resolution

Stereo super-resolution is a technique that utilizes corresponding information from multiple viewpoints to enhance the texture of low-resolution images. In recent years, numerous impressive works have advocated attention mechanisms based on epipolar constraints to boost the performance of stereo super-resolution. However, techniques that exclusively depend on epipolar constraint attention are insufficient to recover realistic and natural textures for heavily corrupted low-resolution images. We noticed that global self-similarity features within the image and across the views can proficiently fix the texture details of low-resolution images that are severely damaged. Therefore, in the current paper, we propose a stereo cross global learnable attention module (SCGLAM), aiming to improve the performance of stereo super-resolution. The experimental outcomes show that our approach outperforms others when dealing with heavily damaged low-resolution images. The relevant code is made available on this link as open source.

1. Introduction

It is widely recognized that, super-resolution is a problem that is inherently ill-posed. Many studies in the field of single-image super-resolution have achieved impressive results by constraining their model’s solution with prior knowledge, as confirmed by the literature [5], [19], [32], [44], [46]. Unlike single-image super-resolution, stereo super-resolution can use inter-view information to reconstruct the subtle details of low-resolution images. In recent times, the need for image resolution has increased greatly due to the advancements in AR/VR and autonomous driving technologies. Thus, stereo super-resolution has gained significant attention among researchers, resulting in a myriad of remarkable results as observed in [2],[30],[34],[38],[40].