CERNet: A Physically Inspired Solution-Space Compression Framework for Low-Light Light Field Enhancement

This paper proposes a physically inspired and interpretable framework for low-light light field (LF) image enhancement, named the Converge-Enhance-Redistribute Network (CERNet). CERNet introduces a solution space compression perspective for low-light LF restoration, which treats all possible LF reconstructions as a set containing pseudo-solutions and removes them by reinforcing physical cues. It leverages multi-view observations to enhance the rank of the implicit observation matrix while encouraging radiometric consistency (i.e., the intensity and directional coherence of ray radiance) for solution space physical plausibility. Unlike previous methods that primarily focus on geometric feature-level fusion or alignment, we promote deterministic enhancement by designing a physics-inspired architecture that guides the solution space toward a physically plausible domain. Specifically, it first constructs the solution space structure by processing LF features in terms of Ray, Illumination, and Photon (RLP), thereby enhancing radiometric consistency, viewpoint discriminability, and cross-view complementarity as the three core elements of solution-space compression. Based on this structure, the framework subsequently performs a Converge-Enhance-Redistribute (CER) process, mimicking the inverse procedure of physical imaging, to further reinforce these core elements and promote a stable approach of the solution space toward a physically plausible domain. Experiments on synthetic and real datasets show that CERNet substantially outperforms state-of-the-art methods in restoration quality and geometric consistency.