Fig. 1: The principle of F-VCD.

a The optical setup of the FLFM system. The inset shows the phase structure of the DOE and its equivalent MLA. b The pipeline of F-VCD construction, including data generation (the upper row) and F-VCD training (the bottom row). In the data generation branch, via simulation, high-resolution stacks are first transformed to light-field views, and then these views are degraded by noise and background quantified from experiments to yield “Noisy LF views”. In the F-VCD training branch, these noisy views are sent to “F-Denoise” network to suppress the severe noise. Finally, with the F-Reconstruction network, high-resolution 3D results can be produced. Among the network inference, each network is optimized by the corresponding loss function (denoise loss for F-Denoise, reconstruction loss for F-Reconstruction). c Real-time inference of 3D images from the recorded 2D light-field images by a trained F-VCD model.