Research On Key Technologies Of Multi-view Light Filed Imaging Based On Array Cameras

Conventional imaging equipment often relies on complex and expensive front-end hardware for long-range high-resolution imaging,and one-shot imaging is only to address a two-dimensional projection of the scene with a certain angle of view.Based on multi-view geometry,computational photography theory and self-developed array cameras,this thesis studies the key techniques of light field imaging such as image super-resolution reconstruction,depth estimation,refocusing,deblurring.The main research contents are as follows:(1)An overall design of array cameras is presented,and a technical system of light filed imaging is constructed,including super-resolution model,process of image super-resolution reconstruction,camera calibration,image registration,depth estimation,image fusion,and image deblurring,etc.(2)The depth estimation and refocusing methods of array cameras are studied.According to the characteristics of parallax depth information and depth of 3D projection information,a depth estimation method based on the combination of parallax method and projection method is presented.An image refocusing algorithm based on Fourier slice principle is discussed.Some real-world scenariosn are photographed by array camera to verify the applicability of this refocusing algorithm.(3)A super-resolution algorithm based on array cameras is studied.Firstly,based on the in-depth study of the alternating direction method of multipliers(ADMM),and through appropriate improvement,a mathematical model of super-resolution robust to noise is established,and a super-resolution algorithm based on ADMM is formed.Then,the maximum common factor solution method and GCD-based super-resolution algorithm are discussed,and an image super-resolution algorithm based on ADMM-GCD and its experimental results are given.(4)A blind deblurring algorithm for array camera images is studied.Firstly,the same scene image acquired by multiple adjacent lenses is registered by using the optical flow algorithm.Secondly,the multiple images after registration are estimated by the multi-images debluring algorithm to estimate the respective blurring kernel.Then,according to the kernel estimation,it uses a non-blind deblurring algorithm to accurately estimate a clear image.In the process of(non-)blind deblurring algorithms,the ADMM algorithm is used to perform fast alternating iterative operations in the frequency domain.Finally,the experimental results show the robustness of the algorithm to noise and its superiority in image recovery efficiency and effectiveness.(5)A parameter-robust balanced image deblurring algorithm based on ADMM is studied.Based on the above work(4),a non-blind deblurring algorithm based on balanced image deconvolution problem and the ADMM algorithm is presented.The experiment results,comparing with some common image deblurring methods,illustrates the robustness and applicability of the algorithm.