Learning-based light field view extrapolation

Abstract

The emergence of light field cameras has challenged the position of traditional cameras in recent years. As light-field cameras become more and more popular, researchers are increasingly studying the light field. However, light field cameras usually compromise in the spatial or angular domain through sparsely sampling, due to the existing tradeoff between the two domain resolution. The most advanced approach is based on machine learning to train the convolutional neural network and gain the high-quality novel views. In this dissertation, I managed to synthesize the novel views by extrapolation, achieved by training deep learning network, based on the most advanced interpolation view synthesis method. There are two components, the disparity prediction network and the color estimation network, that need to be constructed using two sequential CNNs. The two components are trained in MATLAB, through making the error between the synthetic and real images as small as possible. The superior novel views that output by learning-based view extrapolation method are shown in this dissertation. I evaluate the results by showing the measure parameters, PSNR and SSIM, and visual demonstration. In addition, I also analyze the reason of the output novel views that are not of high quality.