Sparsity Analysis Of Light Field And Its Application In Multi-view Data Acquisition

Light field rendering(LFR)is an attractive option to generate novel views because of its low complexity rendering and photorealistic results.It can render novel views from the collected multiview images with simple interpolation operation,and the reconstruction results are good.Therefore,LFR has important research significance and application value in virtual reality,3D imaging,multi-views system and other computational vision fields.However,the major limitation of LFR is its reliance on oversampling to reduce aliasing effects and improve the quality of rendering.Thus,the large number of captured images necessary for alias-free rendering has become an important factor restricting the development and application of LFR technology.Reducing the number of collected images required by LFR,that is,to determine the minimum sampling rate of light field,has become an important problem of LFR technology ensuring the reconstructed views without distortion.To determine the minimum sampling rate for LFR,conventional approaches have applied Fourier theory to obtain spectral support for light fields.Based on the bounds of spectral support,the minimum sampling rate of a light field under certain assumptions(i.e.,a Lambertian surface and nonocclusion)can be determined.However,in reality,there are many scenes characterized by non-Lambertian surfaces,occlusion,and other unknown attributes.Thus,we cannot accurately describe the impact of these scene features on the light field spectrum and determine a reasonable light field sampling rate.Additionally,the light field signals obtained for these scene attributes are not bandlimited.Therefore,the process of light field sampling only relies on the spectrum analysis method will lead to spectrum leakage or unreasonable bandwidth estimation,resulting in the distortion of reconstructed novel views.To above problems,this paper further studies and optimizes the sampling rate of light field using compressed sensing theory.The research work is as follows:(1)The research of light field sampling method is based on the discrete cosine sparse basis.To analyze the sparse characteristics of light field signals,first,we presents a discrete cosine sparse basis(DCSB)to sparse sample light field that relies on the discrete cosine transform(DCT),the relationship of mapping and transformation between camera plane and imaging plane,and the compressed sensing theory.Second,based on the compressed sensing theory,we utilize Bernoulli random measurement matrix to compress the light field to more reasonably reduce the amount of data required for rendering novel views.Finally,sparse Bayesian learning(SBL)is used to obtain the sparse solution of the reconstructed light field based on Bayesian rules.According to the sparse solutions of the light field reconstruction,several groups of experiments are carried out in different real scenes and public dataset scenes.Compared with other classical sampling methods,the reconstructed novel views and its EPI verify the effectiveness of DCSB.(2)The sampling analysis of light field signal is based on the Wavelet Transform theory.To explore the sparse basis which more suitable for light field sampling and optimize light field sampling with simpler numerical analysis,we starts from the sampling theorem and finds the orthogonal wavelet basis which suitable for sparse light field signal through the assumption of sampling interval and bandwidth.We use this orthogonal wavelet basis to transform the light field,so as to study the sparse characteristics and sampling method of the light field.By analyzing the scale function and translation component of wavelet basis,we solve the problems of band limit of signal spectrum and equal interval sampling of signal in traditional sampling methods,and sample a more complete and comprehensive light field signal.Additionally,the nonlinear combined light field signal is linearly represented by its discrete value by orthogonally discretizing the scale function and translation component of the wavelet basis.Finally,the experimental results prove the efficiency of sparse sampling and reconstruction of light field signal using orthogonal wavelet basis.(3)The research of light field sparsity analysis method based on time and frequency domain conversion.Based on Fourier analysis theory,we present a light field sparse analysis(SALF)method.First,according to the influence of the resolution of the rendered view on light field sampling,we analyze the spectral support of camera plane through a spectrum analysis of the imaging plane.Based on the spectral support of the camera plane,the capture interval of cameras can be reasonably determined.Second,we sample a light field in the frequency domain and use the Fourier projection-slice theorem to simplify the expression of light field sampling.Finally,we use a voting scheme to select camera positions in which the frequency coefficients are nonzero.The experimental results show that our approach can optimize the sampling rate in LFR using these camera positions.These experiment results highlight the improved rendering quality of the proposed strategy,which outperforms those of other compared methods.