Light-field Image Super-resolution Using Convolutional Neural Network

By changing the optical structure of the ordinary camera,light field cameras can capture the light's direction information while capturing the intensity information.Using the angular information captured by the light field camera,combined with computational imaging techniques,light field imaging has made remarkable progress in super-resolution reconstruction,three-dimensional reconstruction,depth estimation,target detection,face recognition and other fields.Because the pixels on the image detector are fixed,the increase of the angular resolution will inevitably lead to the decrease of the spatial resolution,which limits the application of light field imaging.For the problem of low resolution of light field image,this paper explores the light field image super-resolution algorithm based on convolutional neural network.The main work is as follows:1.We summarize the background and research status of the light field camera and super-resolution reconstruction technology,introduce the basic theory of Light field photography and super-resolution reconstruction,including parallax-based depth estimation and focus-based depth estimation,then introduce the traditional method of Light-Field image Super-Resolution.2.We describe two kinds of light field image rendering methods and convolutional neural network theory,then implement a method for Light-Field image angular Super-Resolution using Convolutional Neural Network,the operations consists of preprocessing,feature extraction,non-linear mapping,reconstruction and losses computation.Experimental results show that Convolutional Neural Network can maintain details of the new perspective image after reconstruction,but has poor performance in a large range of depths.3.On the basis of convolutional neural network,we present Perceptual Losses for Light-Field image angular Super-Resolution.Combining the perceptual loss applied in the field of image recognition with the per-pixel loss,we present a method for Light-Field image Super-Resolution using Convolutional Neural Network.Experimental results show that the super-resolution reconstruction algorithm based on perceptual loss can not only improve the angular resolution of the light field images,but also maintain more details of the new perspective image after reconstruction.