DistgEPIT: Enhanced Disparity Learning for Light Field Image Super-Resolution

Kai Jin, Angulia Yang, Zeqiang Wei, Sha Guo, Mingzhi Gao, Xiuzhuang Zhou; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 1373-1383

Light Field (LF) cameras capture rich information in 4D LF images by recording both intensity and angular directions, making it crucial to learn the inherent spatial-angular correlation in low-resolution (LR) images for superior results. Despite impressive progress made by several CNN-based deep methods and pioneering Transformer-based methods for LF image super resolution (SR), most of them fail to fully leverage the LF spatial-angular correlation and tend to perform poorly in scenes with varying disparities. In this paper, we propose a hybrid method called DistgEPIT that implements an enhanced disparity learning mechanism with both convolution-based and transformer-based modules. It enables the capture of angular correlation, refinement of adjacent disparities, and extraction of essential spatial features. Additionally, we introduce a Position-Sensitive Windowing (PSW) strategy to maintain consistency of disparity between the training and inference stages, which yields an average PSNR gain of 0.2 dB by replacing the traditional padding and windowing method. Our extensive experimental comparison with necessary ablation studies demonstrates the effectiveness of our proposed method, which ranked 1st place in the NITRE2023 LF image SR challenge.