Research On Depth Estimation Algorithm Of Light Field Images Based On Stereo Matching

Light field cameras capture the direction and the angle of the rays simultaneously,extending the traditional two-dimensional image to four dimensions,which is more creatable.The research in the field of computer vision around the light field images has attracted wide attention.In which,depth estimation,a branch of computer vision,is of far-reaching significance and has extensive application scenarios.Method of depth information acquisition based on light field images is passive,which fully utilize the excellent characteristics of images obtained by light field cameras.This method also avoids the contradiction between monocular and multi-ocular depth estimation.The stereo matching algorithm can match the corresponding pixels of light field images to obtain the disparity of each pixel,and then obtain the depth information.However,the traditional stereo matching algorithms have some problems when calculating disparity of each pixel,such as low matching precision,inaccurate depth estimation,discontinuous and unsmooth depth under complex occlusions.Therefore,an improved stereo matching algorithm for depth estimation of light field images is proposed in this paper,and a method for handling multi-occlusion problems in depth estimation of light field is proposed.Based on the above analysis,the main research contents and innovations of this paper are as follows:(1)An improved stereo matching algorithm for depth estimation of light field images is proposed.First,by analyzing the characteristics of sub-aperture images in light field,different labels are set to displace.Second,matching cost volume is established based on variance and gradient error,and the guided filtering is used for keeping the edges.Third,the label corresponding the minimum cost is taken as the initial result of disparity estimation.Finally,constant time weighted median filtering is further used for optimization to make the final depth more clear and accurate.(2)The method to process occlusions in light field images is proposed.Based on the classification of occlusion problems in light field images,the occlusion problems are modeling.With the principle of pixel consistency in spatial domain and angle domain,optical flow algorithm and clustering algorithm are used to deal with occlusion problems.The result shows the algorithm can effectively extract the occlusion edges and get accurate clustering results.(3)A method for handling multi-occlusion in depth estimation of light field is proposed.Based on the above methods of processing occlusions in light field and the improved matching cost calculated by variance and gradient error,the cost volume is established for estimate the initial depth map,then Markov Random Field and constant time weighted median filtering are used to optimize the final depth map,aiming to get more accurate depth estimation results in case of occlusions.The experimental results show that the method studied in this paper has good performance both subjectively and objectively.Subjectively,clear depth maps with continuous and sharp edges can be observed.Objectively,the evaluation index of average MSE is lower than other comparative algorithms.