Research On Synthetic Aperture Imaging Based On Camera Arrays

Light field(LF)cameras have drawn much attention in recent years since they can record 3D information of a scene,and enable applications such as refocusing,depth sensing and superresolution(SR).Compared to commercial microlens array based LF cameras,camera arrays have a much brighter prospective with the development of computing capacity and camera miniaturization.However,the baseline of a camera array is relatively wider than that of a plenoptic camera,which leads to sparse samplings in angular direction and limits the performance of many LF imaging algorithms.Accordingly,this thesis presents an extensive theoretic and technical investigation on synthetic aperture imaging based on camera arrays.First,a depth estimation method is proposed to recover the depth of a scene captured by a camera array.The reliability is calculated with the initial depth estimation,and further used in global optimization as guidance.Our method achieves superior performance compared with other state-of-the-art methods,especially in textureless regions.Second,a novel synthetic aperture imaging method is proposed to improve the image resolution and bokeh performance of images captured by a camera array.The depth is firstly estimated to recognize the focused and unfocused regions.A bokeh rendering approach is then presented to render an esthetical bokeh in unfocused regions.The bokeh image is further used as a regularization term in a reconstructed-based SR framework.Third,an SWT-based multi-focus image fusion method is proposed to fuse the LF refocused images.Compared to traditional DWT-based fusion method,our method avoid image down-sampling by using interpolation to wavelets,which helps to alleviate artifacts brought by down-sampling.Combined with the proposed synthetic aperture imaging methods,any selected depth ranges can be arbitrarily refocused.Forth,we analyze the issues which determine the quality of SR reconstruction and investigate the influence of decent direction,step size,and number of iterations to the convergence of SR process.We infer the approximate analytical expression of optimal step size,and propose an adaptive line search strategy to improve the convergence of SR reconstruction.We have conducted extensive experiments on several public datasets to fully demonstrate the effectiveness and superiority of the proposed methods.Moreover,we have developed our own LF imaging devices,and conducted experiments in our laboratory to test the practicality of the proposed methods.