Study On Super-resolution Reconstruction Of Light Field

Light field camera is a new multi-view image capture equipment.A light-field camera is able to acquire 2D spatial information and 2D angular information of a scene in a single shot by adding a micro-lens array between the main lens and the pixel sensor which has incomparable advantages compared with other traditional cameras.Aiming at the problem of low spatial and angular resolution of light field camera and combining with the advantages of Convolutional Neural Network(CNN)in image processing,this thesis is focused on super-resolution algorithm of light-field image respectively for occlusion scenes and high light scenes.The main work of this thesis is listed as follows:(1)We summarize the research situation of traditional 2D image and light field super-resolution reconstruction technique,then introduce the basic theory of light-field imaging,the acquisition method of light-field and its parameter representation,then elaborate the theory of light-field rendering,multi-view images,epipolar plane image and super-resolution reconstruction technique.(2)We study and realize the method of light field super-resolution reconstruction with global and local views.It adaptively chooses local views according to the position of the novel view,and then fuses global and local views as the input of CNN to produce a novel view image.The experimental results demonstrate that the reconstruction effect of this method is not only able to improve spatial and angular resolution of multi-view images,but also to maintain edge information of the novel view,thus effectively avoiding the loss of objects and yielding a better reconstruction effect.(3)We take advantage of the clear texture structure of the EPI in the light field data and model the problem of light field reconstruction from a sparse set of views as a CNN-based angular detail restoration on EPI.A deep residual CNN is used to restore the angle details of the EPI image in the sparse input angle of view,and the results are used to obtain the final novel views.This method can reconstruct novel views from high light scenes.The experimental results demonstrate that the reconstruction result of the proposed method is superior to the state-of-the-art methods.And in peak signal to noise ratio get 0.3 dB promotion and structural similarity image measure upgrade about 0.2.