Light Field Reconstruction Using Infinite-dimensional Compressed Sensing

The purpose of imaging is to get the interaction between light field and various substances in the physical world.The traditional camera only could capture the projection of light field on a focal plane and discard the angular signal.In recent years,the light field imaging,which is a new type of imaging system that is capable of collecting the entire high dimensional data,is attractive for researchers.Light field cameras could be applied to the correlation application domains,such as scene focusing,depth estimation and three-dimensional display.Additionally,light field camera mainly includes the equipment based on multi-camera array and based on microlenses array.Nevertheless,the light field acquisition equipment based on multi-cameras has considerable size and is so expensive.In the same time,the single camera with an array of microlenses has spatio-angular resolution tradeoff.Thus,researching the sampling and super-resolution of light field has important sense.Actually,the light field image is a classical sparse signal that could be sampled at a low rate.Compressed sensing(CS)is the sampling method,which compresses the signal in the sampling step.Moreover,basis of this method is the sparsity of signals.Recently,the light field imaging method based on CS theory has become one of the focuses in the research field.However,the traditional methods for light compressed imaging had suffered from incomplete sparse representation,sparse dictionary with poor generalization ability and window effect and so on.To address these problems,this thesis analyzes the light field structure sparsity and proposes a reconstruction method for light field angular super-resolution based on infinite-dimensional compressed sensing theory.In this thesis,we study the relationship between CS theory and sparse light field,and illustrate the possibility of using CS technology to reduce the number of light field samples.Subsequently,we analyze the sparse property of light field in shearlet domain,and study the feasibility of applying CS to recovery light field by studying the reconstruction method using shearlet transform.Moreover,we introduce an idea for continuous Fourier spectrum recovery and study light field reconstruction using sparse Fourier transform.And the simulation results illustrate that the optimization for continuous spectrum supports would alleviate window effect.Afterward,according to epipolar-plane image theory,structured sparse prior in light field is proven.What is more,we introduce light field structure sparsity and present a light field reconstruction algorithm based on the structure sparsity by improving a traditional CS method.And the experimental results demonstrate that compared with the traditional CS method,our algorithm has the same performance with fewer measurements.On this foundation,we introduce infinite-dimensional approximation and propose a light field reconstruction algorithm based on infinite-dimensional compressed sensing.Finally,the experiments reveal that the peak signal to noise ratio(PSNR)of our reconstruction algorithm is higher 3dB than those traditional light field recovery methods for Lambertian scene.Moreover,for non-Lambertian scene,our algorithm has the same performance.