Study Of Depth Estimation Based On Refocused Image And EPI Image Of Light Field

The continuous development of light field rendering theory and the continuous evolution of the plenoptic function have laid a theoretical foundation for the acquisition of light field.The microlens light field camera and the array light field camera designed based on the biplane parameter representation model of the light field can realize the simultaneous acquisition of spatial light position information and angle information.Compared with the traditional two-dimensional image,the 4D light field image that simultaneously records the spatial light position information and the angle information has a greater advantage in the scene depth.This article is fully studying the light field digital refocusing and the light field sub-aperture image stereo matching technology.On the basis of,research on depth estimation technology based on light field multi-focus image and EPI image,the specific research work and innovation are as follows:1.An acceleration algorithm for light field depth estimation based on digital refocusing in frequency domain is proposed.Defocus combined with correlation evaluation depth estimation algorithm is based on the calculation of multiple light field digital refocusing images.Traditional spatial digital refocusing involves a four-dimensional spatial transformation and integration process,which seriously affects the calculation speed of refocusing images.In this paper,the frequency domain digital refocusing technology is introduced to transform the spatial four-dimensional space transformation into the frequency domain to take two-dimensional slices,which improves the speed of refocusing images.In addition,in the process of taking two-dimensional Fourier slices in the four-dimensional frequency domain space,the frequency domain coordinate conversion matrix is optimized,so that the slice operation is reduced from 16-point interpolation to 4-point interpolation,which further reduces the speed of refocusing image acquisition.Which ultimately greatly improves the speed of depth estimation.The experimental results show that the introduction of the frequency domain refocusing algorithm can shorten the refocusing time by 24.4seconds/frame on average,and the optimization of the coordinate transformation matrix can shorten the refocusing time by 76.6 seconds/frame on average,and after the refocusing algorithm is accelerated,the quality of the final acquired depth map is not significantly reduced.2.A light field image depth estimation algorithm based on improved EPI-Net is proposed.Aiming at the problems of large amount of network parameters,underutilization of features,low quality and small size of the acquired depth map when traditional EPI-Net acquires depth maps,a light field image depth estimation framework based on improved EPI-Net is proposed,which uses Dense-Block to extract the high-level features of the image,which solves the problem of large parameters,while making the network feature map fully applied,thereby improving the quality of the depth map.The experiment verifies the rationality and effectiveness of the proposed algorithm.Compared with the traditional EPI-Net light field depth estimation network,the depth map obtained by the improved network has clearer texture features visually.Quantitative evaluation shows that the mean square error and bad pixel rate of the depth map obtained by the improved network decrease on average by 29.07%and 32.86%.