Boosting Light Field Image Super-Resolution via Manually Shearing Disparity

Light field imaging, with its ability to capture both the intensity and directionality of light rays, has gained significant attention in various fields including virtual reality and computational photography. Super-resolution (SR), the process of enhancing image resolution, is essential for improving the visual quality and detail of light field images. In this paper, we integrate data augmentation techniques into our approach to further enhance the generalization capability of existing light field image SR models. By manually adjusting the disparity range of existing light field image datasets, we augment the training data and introduce variations that help the models learn robust features and reduce overfitting. By doing so, we tame the models to capture larger contextual information among the sub-aperture images of light filed and improve their SR performance without the need for extra datasets. We demonstrate the effectiveness of our method through extensive experiments on benchmark datasets, showcasing existing light field SR networks to achieve superior reconstruction performance and angular consistency.