Restoring extremely low-light light fields via decomposition-enhancement and frequency-pixel fusion

Light field (LF) technology records both spatial and angular information of light rays in space, yet it faces significant challenges under extreme low-light conditions. The simultaneous processing of the complex degradations of low-light LF restoration suffers considerable difficulties. To address this issue, we propose what we believe to be a novel decomposition-enhancement method that leverages properties of frequency space, allowing for the partial decomposition of luminance and noise, to achieve frequency-pixel domain information fusion. Our approach decouples the low-light enhancement task into three sub-tasks: brightness adjustment, noise suppression, and detail refinement, each tackled independently. In the frequency stage, we employ two distinct branches: one dedicated to brightness adjustment by estimating a transformation map, and the other focuses on noise suppression through the reconstructing clear phase map. During the fusion stage, we implement a cross-attention module to effectively integrate information from the frequency-enhanced outputs and the original low-light input for the purpose of feature enhancement. Finally, in the pixel stage, we utilize a spatial-angular attention module to recover fine details by exploiting global information within each view and complementary information among all views. Extensive experiments demonstrate that our method achieves competitive performance compared to state-of-the-art methods, showcasing its effectiveness in enhancing light field images under challenging low-light conditions.