Research On Refined Depth Estimation Algorithm From Light Field

With the rapid development of artificial intelligence technology,as an important component of artificial intelligence technology,the application prospects of computer vision are very extensive.As an important part of computer vision technology,depth estimation algorithm of light field has broad application scenarios and great research value.However,there are some disadvantages in traditional depth estimation techniques.For example,the multi-view system requires a large amount of equipment and relatively high cost.However,the monocular vision system has too few viewing angles,so with low accuracy.The appearance of the light field camera perfectly solves the problems in the monocular and multi-view system vision systems,and injects fresh blood into depth estimation algorithm.In the light field camera system,a microlens array is placed between the main lens and the image sensor,the amount and direction of each incident light are simultaneously recorded,which is equivalent to simultaneously collecting the spatial and angular information of the object.Therefore,the rich information carried by the light field image is very beneficial for the subsequent depth estimation.However,the baseline of the light field image is very narrow,the spatial resolution and angular resolution are difficult to compromise,and there are some other problems such as occlusion,etc.The calculation results of traditional dense depth estimation algorithm are not very accurate,and because of the large amount of data,the calculating efficiency is low.Therefore,studying the depth estimation algorithm suitable for light field,especially for sparse light field is the focus of this paper.Based on the above analysis,this paper mainly studies the ultra-fine depth estimation algorithm of light field.The main research contents and innovations are as follows:1.The related principles of depth estimation algorithm of light field are explained,including the two-plane parametric model of the light field,the EPI visualization theory;the principle of light field camera collecting light,the principle of light field refocusing;and using the Matlab toolbox to preprocess the original image captured by light field camera.2.The basic framework of the depth estimation algorithm is studied.The main process includes the preprocessing of the input image,the extraction of the initial depth of light field,the refinement of the depth,and finally the ultra-fine depth map will be achieved.Four initial depth estimation algorithms are analyzed:multi-view stereo matching method,EPI-based method,focus stack method,deep learning-based method.On this basis,the refinement processing method of TV(total variation)regularization and guided filter are studied.3.A depth estimation algorithm based on EPI and TV regularization is proposed.The initial epipolar analysis is confined in a small range to improve the efficiency of the algorithm,combining with a regression test to reduce estimation error and a weighted summation to improve accuracy.Finally,the initial depth of field is refined with the TV regularization scheme.The experimental results show that the algorithm extracts the depth of field of dense light field very accurately and has good adaptability to sparse light field.4.A depth estimation algorithm based on convolutional neural network(CNN)is proposed.Based on the EPI method,a multi-flow network corresponding to four viewpoint directions(horizontal,vertical,left and right diagonal directions)is designed,and the features of these four directions are extracted separately and then merged.Aiming at the sparse field,a special data amplification method for central viewpoint shift is proposed.Finally,using the guided filtering and TV regularization methods,the depth extracted by the CNN is further refined.The experimental results show that the proposed algorithm can not only extract the ultra-fine depth of field of computer synthesized light field,but also adapt to the real light field.It can not only obtain the ultra-fine depth of dense light field,but also obtain the ultra-fine depth of sparse light field.