Frequency-Guided Multi-View Super-Resolution for Light Field Images

Abstract

Light field imaging captures both spatial and angular information but suffers from an inherent trade-off that limits per-view spatial resolution. This work addresses center-view super-resolution (SR) from multi-view inputs and reveals that naive view aggregation fails to effectively exploit angular information, even as the number of views increases. To overcome this limitation, we propose an FFT-guided residual multi-view SR framework that performs structured reconstruction in the frequency domain. Extensive experiments on real light field data show that the proposed method consistently outperforms bicubic, single-view CNN, and naive multi-view baselines, achieving up to 40.58 dB PSNR and 0.9600 SSIM. Furthermore, compact angular subsets, such as 9 views, outperform larger ones, highlighting that effective fusion is more critical than view count. These results demonstrate that frequency-domain residual learning provides a principled and robust approach for multi-view light field super-resolution.