Research On Image Super-Resolution Algorithms Based On Deep Learning

The spatial resolution of the image determines how much information is included in the image.And it has been the focus of image science research how to improve the image resolution.Image super-resolution reconstruction technology aims to generate a high-resolution image from one or more low-resolution images with machine learning or computer vision knowledge.In recent years,deep learning in image classification,face recognition and other tasks have achieved great success.This paper also tries to apply deep learning theory into image super-resolution task and explore super-resolution algorithms based on deep learning and its application.In this paper,we first summarize the research status of super-resolution technology,and then propose several super-resolution reconstruction methods based on deep learning and its application in face hallucination.First,we propose a super-resolution model based on deep encoder-decoder symmetrical network.The fully convolutional network(also known as encoder-decoder network)has recently achieved better results in image segmentation and image denoising due to the deconvolution operation.In previous network,the image details may be lost with the increasing layer of network,which is not useful for image recovery tasks.Therefore,we try to build a symmetrical network consists of convolutional and deconvolutional layers to solve the super-resolution task.The experimental results verify the effectiveness of this method.Then,a super-resolution model based on multi-scale deep symmetrical network is raised.Multi-scale analysis has always been the guiding principle of image processing technology.It’s a new attempt to use multi-scale deep network to simulate multi-scale analysis process.Generally,the deeper the network is,the more abstract feature will be learned in deep learning,which is more conducive to follow-up tasks.We build a multi-scale deep network,which has achieved better reconstruction effect.In addition,we introduce the phase congruency theory to extract the image edge,which regularizes the ill-conditioned problem of super-resolution to a certain extent.At last,we discuss face hallucination based on deep learning theory.Image prior knowledge plays an important role in improving the performance of the most visual tasks.And low-resolution image is introduced in back projection algorithm.Therefore,we combine iterative back projection with convolutional neural network taking the form of network.In addition,we introduce generative adversarial network into face hallucination.Training two networks together make the reconstructed image more close to the original image.The experimental results show the effectiveness of adversarial super-resolution model.