An Effective Occlusion Edge Prediction Method in Light Field Depth Estimation

Occlusion is one of the most complicated and difficult problem to deal with in light field depth estimation. Previous methods usually use edge detection of central view as the basis for subsequent depth estimation. However, few algorithms take the correlation between light field sub-aperture images into consideration, which makes the incorrect depth estimation in the occlusion areas and object boundaries. In this paper, based on the refocusing theory, we explore three occlusion prediction methods and derive model to make full use of the correlation between different sub-aperture to handle the occlusions in depth estimation, the obtained occlusion map is used as a guidance for the depth estimation. Markov Random Field is built to regularize the depth map and smoothing filtering, which is used at the background area of the non-occlusion area to process noise in order to improve estimation results. Experimental results on the real light-field data sets show that the proposed algorithm can greatly improve the depth estimation results compared with state-of-the-art algorithms, especially in complex occlusion areas.