Light Field Depth Estimation And Super-resolution Reconstruction In Complex Occlusion Scenes

The light field camera can capture the space and angle information of the scene at the same time,so there are advantages of multidimensional data in deep estimation which also makes the light field very popular in numerous visual fields..However,the depth estimate of the occlusion area is low enough to meet the reconstruction requirements.Due to the limit of hardware conditions,the light field camera cannot simultaneously increase spatial resolution and angular resolution,so the sacrifice part of the spatial resolution is selected to record the scene multi-view information,which makes the image quality of the light field camera decreased.In order to solve the problems mentioned above,this paper proposes an anti-Occlusion depth estimation algorithm and an super-resolution reconstruction algorithm based on reclassification.This article is summarized as follows:(1)Research on occluded areas in complex scenes.First,this paper constructs an occlusion detection model based on the difference between the sub-aperture images of the light field,and then the accurate occlusion detection map is obtained based on the occlusion detection model,and the clustering algorithm is used to divide each occlusion point window into the occlusion area and In the non-occluded area,the occluded edge is obtained.Finally,multiple features such as the occlusion detection map,the occlusion edge,and the center perspective map are combined and optimized using the MRF model to improve the accuracy of the depth estimation results.(2)A new type of optical field depth estimation network is proposed.The network uses the occlusion detection map as a new network input feature,and combines horizontal and vertical epipolar plane images(EPI)to build a three-input branch network,and then uses the EPI index as the time dimension in the branch network and uses 3D convolution The kernel replaces the 2D convolution kernel,and then connects the branch network to complete the depth estimation of the full convolution.A super-resolution reconstruction model of light field based on full focus is constructed to improve image resolution.The model first fuses the image degradation model,bokeh rendering and depth map to select a single focus super-resolution reconstruction,nd then uses the focus stack to perform super-resolution reconstruction of all-focus.Finally,design experiments from multiple angles to verify the superiority of the algorithm both qualitatively and quantitatively.