Efficient Low-Light Light Field Enhancement With Progressive Feature Interaction

Light field (LF) imaging can be degraded under low-light conditions, which brings difficulties in scene sensing, understanding, and downstream applications. To address this issue, low-light LF enhancement has been introduced. However, existing methods still suffer from high computational costs and sub-optimal performance due to ineffective global-local stacks modeling and underutilized inter-view correlations. In this paper, we propose a novel approach, named
PFInet
, for efficient low-light LF enhancement. Specifically, we propose to first incorporate global and local dependencies at multiple scales and then aggregate these representations for a richer spatial context understanding. Moreover, we develop a progressive feature interaction (PFI) strategy, which sequentially performs interaction on multiple subspace-specific features to exploit the inter-view correlations of LF images. Extensive experimental results on benchmark datasets demonstrate the superiority of our method, surpassing the second-best method by 0.38 dB on average in terms of PSNR. Furthermore, our method achieves efficient low-light LF enhancement, as evidenced by a good tradeoff between the inference time and reconstruction performance.