Texture Complexity Adaptive Light Field Image Super-Resolution
Qianyu Chen, Shuanglin Wu, Jungang Yang, Wei An · 2024
This paper aims at reducing the computational cost of light field (LF) image super resolution (SR) techniques while preserving competitive performance. Inspired by the recent advance in designing sub-networks for different texture patches in single image SR, we propose a light weight LF image SR framework. Specifically, the input LF image patches are classified into different groups for reconstruction based on the texture complexity evaluations. Each image group corresponds to a careful-designed subnet that explores the latent correlations of LF in spatial, angular and epipolar dimensions. As the smooth area can be well super-resolved with less prior knowledge than the complex one, the parameter demand of feature extraction subnet for smooth patches can be much smaller. We also introduce a weight-sharing strategy for each SR subnet to avoid parameter redundancy. Experiments show that the proposed method can achieve reliable LF image SR results with lower computation demand.