Single-Exposure HDR Light Field Imaging with Spatial-Angular Collaborative Learning

This paper proposes a novel spatial-angular collaborative learning based high dynamic range light field image (HDR-LFI) reconstruction method to solve the problem of high dynamic range (HDR) light field imaging with a single exposure in light field camera. The proposed method leverages a dual-network architecture, i.e., spatial restoration network (SRN) and spatial-angular collaborative reconstruction network (SACRN). SRN is designed to focus on restoring LFI's over-exposure regions, while SACRN focuses on projecting the standard dynamic range image to the HDR domain while further enhancing over-exposure regions. SACRN utilizes Fourier convolution to learn spatial information, and jointly learns local and global information in the spatial-angular dimension using spatial-angular convolutions and non-local blocks. Finally, realistic HDR-LFIs are reconstructed while maintaining angular consistency. The experimental results show that compared to the representative methods, the proposed method generates higher quality HDR-LFIs.