Fast Light Field Reconstruction Using Convolutional Neural Network on Lenslet Images

Light field (LF) imaging is superior to traditional photography in terms of the quantity of information it can capture. It collects angular and spatial data from all directions. Despite this, its poor resolution makes it impossible to take advantage of this immense potential. We attempted to mitigate the effect of this disadvantage in our research by using a deep learning technique. To facilitate our task, we have converted multiple images of the same scene from a 4D variant into a single 2D image (lenslet image). Thus, using very deep models to recover lost details and reconstruct LF images is no longer necessary. We trained our model for one reconstruction task, but the same model can be retrained for other reconstruction tasks. Our results show that the proposed method reconstructs high-quality images faster than other approaches.