Efficient Light Field Image Super-Resolution via Progressive Disentangling

Gaosheng Liu, Huanjing Yue, Jingyu Yang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 6277-6286

The performance of light field (LF) image super-resolution (SR) has been significantly improved with the development of deep learning techniques. In recent state-of-the-art methods increasingly deeper and wider networks with a massive number of layers are employed to improve SR performance. However these approaches often incur heavy computational costs hindering efficient inference and practical applications. In this paper we address the problem by introducing an efficient network for LF image SR. Specifically we propose an efficient progressive disentangling block (PDistgB) where the intermediate LF feature is progressively channel-wise split and selectively domain-specific disentangled. The PDistgB can well incorporate the LF structure prior while requiring fewer computational costs compared with existing disentangling strategies. In addition we apply Transformer on the angular domain to incorporate angular correlations from all views for further improving the SR accuracy. Experimental results on public datasets demonstrate that our method achieves state-of-the-art performance with high efficiency. Codes and models are available at https://github.com/GaoshengLiu/PDistgNet.