Adapting the Learning Models of Single Image Super-Resolution Into Light-Field Imaging

The emergence of Light Field (LF) cameras has made LF imaging a popular technology in computational photography. However, the spatial resolution of these micro-lens-based LF cameras is limited due to the combination of spatial and angular information, which is the primary obstacle for other applications of LF cameras. To explore the potential of LF imaging, Light-Field Super-Resolution (LFSR) algorithms have been developed to exploit the spatial and angular information present in LF imaging. In this paper, we propose an alternative approach to achieve LFSR using pre-trained Single Image Super-Resolution (SISR) models. We introduce an LF domain-specific adaptation module that can be included in any SISR model to make it suitable for the LF domain. We experimentally demonstrate that three different kinds of SISR models, namely the bicubic degradation handling model, the blur kernel based model, and adversarially trained SISR models for perceptual super-resolution, can be converted to corresponding LFSR models. Our experimental results show that by using a recent state-of-the-art SISR model, we can outperform recently reported LFSR-specific models for bicubic degradation by a considerable margin in both the standard test dataset and the recent NTIRE 2023 LFSR challenge test dataset. In the case of models that handle blur kernel, we observe a significant performance improvement after adaptation. Adversarially trained SISR models also show promising results, with less distortion and better perceptual quality in LF images.