Research On Anti-Occlusion Depth Estimation Algorithm Of Light Field Images

At present,artificial intelligence technology develops rapidly,as an important branch of which,computer vision technology has already reached the industrial level in many practical problems.As a branch of computer vision technology,depth estimation has profound research background and extensive application scenarios,and has always been the focus of the research.However,there are some deficiencies in traditional depth estimation techniques.For example,single-view vision system has relatively less ac-curacy on account of less perspectives,while multi-view vision system requires a lot of equipment and high cost.In recent years,the emergence of light field camera has made up for the contradic-tion between single-view vision system and multi-view vision system,and has triggered a new research boom in depth estimation.The light field camera places a microlens ar-ray between the main lens and the photoelectric sensor to record the amount of light and the direction of incidence for each light,and achieves simultaneous sampling of both spatial and angular information.Therefore,the rich information contained in the light field image is conducive to various computer vision tasks such as depth estimation.However,in a complex occlusion scene,the assumption of angular pixel consistency used in traditional depth estimation on light field images no longer holds and the proba-bility of erroneous depth estimation increases,which leading to fuzzy or distorted depth estimation results.Therefore,the anti-occlusion handling in depth estimation of light field images is the difficulties and focuses of research.Based on the above analysis,this paper mainly deals with the occlusion problem in depth estimation of light field images.The main research contents and innovations are as follows:1.A light-field image occlusion edge pixels extraction algorithm is proposed.In light field image depth estimation,the extraction of occlusion edge pixels is the first and important step of the algorithm.The Canny detector is a common method in this field,but its detection results will include internal textural pixels,it is inaccurate and re-duces algorithm performance.This paper makes full use of the perspective distribution characteristics of the light field image,and uses the optical flow algorithm commonly used in the video field to extract the occlusion edge pixels.The experimental results show that compared with Canny,the optical flow algorithm can extract more accurate occlusion edge pixels of light field image.2.An anti-occlusion depth estimation algorithm of light field images is proposed.In actual scenes,intricate occlusion is a very common phenomenon,which can cause inaccurate depth estimation results in the occlusion area and influence the overall ac-curacy of the algorithm.This paper analyzes the imaging system of light field camera,proposes a complex occlusion model,and obtains the corresponding relationship be-tween space and angle.That is to say,the spatial patch of the occlusion edge pixel and the angular patch refocused to the correct depth have correspondence.Based on the above model and theory,the partitioning result of the spatial patch is used to divide the corresponding angular patch,and depth estimation is performed in multiple regions of the angular patch to resist the influence of occlusion.Experimental results show that the anti-occlusion algorithm proposed in this paper can obtain more accurate and clear depth estimation results in the occlusion area.3.An improved matching cost design algorithm in depth estimation of light field images is proposed.In this paper,based on the calculation of the matching cost by vari-ance,the improved variance of measuring the consistency between the angular pixels and the central view pixel is added,and the weights are adaptively weighted to obtain the final matching cost.The initial depth estimation results are obtained by the matching cost,and then the Markov random field is used to optimize the depth results.Experi-mental results show that the matching cost and the optimization algorithm designed in this paper are helpful to improve the accuracy of depth estimation...