Figure 4.

Illustration of deconvolution algorithms for LFM. (A) Model-driven methods explicitly form the optical transfer matrix H from PSFs and inversely solve the volume vector g from the light field measurement f. Due to the large size of the matrix H, iterative methods such as Richardson-Lucy (Richardson, 1972; Lucy, 1974) are usually adopted (Broxton et al., 2013). (B) Data-driven methods implicitly model the measurement matrix H by training a neural network to learn the mapping relationship between the light field images and the target volumes from the large light field-volume pairs dataset. The mean-square error (MSE) is usually adopted to provide the loss between the predicted volume and the ground truth volume. The loss then backpropagates to update the neural network weights during the training stage. (A) adapted with permission from Broxton et al. (2013) ©The Optical Society.