Fast Disparity Compute Based On Light-field Image And Parallelization Optimization

As a new type of image imaging and description method,light-field imaging technology has developed rapidly in recent years.Co mpared with traditional camera,light-field camera can collect light f rom different directions,and the micro-lens array in light-field came ra can retain the space and angle information of the collected light.The light-field image provides great convenience for computer visi on tasks such as depth estimation due to the rich information it co ntains.However,due to the special structure of the light-field imag e,such as micro-baselines,it also faces new challenges to the dept h estimation task.The main tasks are as follows:(1)A real-time SGM(Semi-Global Matching)depth estimation algorithm for practical application scenarios is proposed.Based on t he analysis of the principle and related characteristics of light-field imaging technology,a customized cost calculation function of stereo matching is designed,and in order to improve the parallel granula rity,the cost aggregation based on Semi-Global optimization is intro duced.(2)A parallel optimization scheme for light-field SGM depth e stimation algorithm based on GPU is proposed and designed,which can fully mine the hardware performance and ensure the real-time p erformance of the algorithm.(3)An iterative optimization algorithm for depth estimation of light-field based on PatchMatch is proposed.In this paper,accordin g to the requirement of high-precision depth map in practical applic ation scenarios,a super-pixel segmentation algorithm sensitive to de pth and edge and a cost calculation function based on multi-view o f light-field are designed.Based on the idea of PatchMatch,the init ial depth estimation results are further optimized iteratively.(4)This paper proposes a spatial propagation strategy based on clustering,and designs a combination feature of color information,spatial information and depth information for super-pixel blocks.Ba sed on clustering algorithm,similar super-pixel blocks are clustered to the same propagation unit to optimize the spatial propagation str ategy of the original PatchMatch algorithm.In summary,this paper studies the light-field depth sensing sys tem for practical application scenarios.According to the real-time a nd high-accuracy application requirements in real applications,this p aper designs a real-time light-field SGM depth estimation algorithm and an iterative optimization algorithm for light-field depth estimatio n based on PatchMatch.Experiments show that the algorithm propo sed in this paper outputs the light-field depth estimation result at 20fps,and the accuracy of the initial depth estimation result is greatl y improved after iterative optimization by PatchMatch.