Image Super Resolution Reconstruction Based On Residual Learning

Image Super Resolution(SR)is an important branch of image restoration.It is widely used in satellite remote sensing imaging,public surveillance video recognition,military detection,biomedical imaging,and multimedia video imaging.And it has a certain application prospect and practical significance.Image super-resolution reconstruction technology refers to using a digital signal processing method to reconstruct a high-resolution image with high-frequency information by one or more low-resolution observation images of the same scene processed by software.In recent years,deep learning,with its powerful learning ability and efficient feature expression capability,has become the most sought after field of research in machine learning.From low-resolution features to high-resolution features,image detail information is extracted layer by layer,and at the same time,deep learning brings new ideas for image super-resolution reconstruction,eliminating the need to manually extract image features from traditional methods.It makes the deep learning model show prominent advantages in image super-resolution reconstruction.This thesis is inspired by the deep learning model and uses a deep residual network model to achieve super-resolution reconstruction of images by using multi-level shortcut links.Based on the current research status,this thesis mainly carries out the following work:1.We summarize the current research status of image super-resolution reconstruction,and we classify and summarize the existing super-resolution algorithms,and then expound its related theoretical basis.2.We discuss the theoretical and structural models of convolutional neural networks in depth,and introduce image super-resolution reconstruction algorithms based on convolutional neural networks,then we point out their research results,which will lay the foundation for subsequent research.Furthermore,we explain the current popular deep residual network model,then analyze the application of residual learning in image super-resolution reconstruction,and prove that the model is superior to the traditional super-resolution reconstruction algorithm.3.In order to improve the visual effect of low-resolution images and increase the amount of detail information of images,we propose a novel residual network structure-Deep multilevel residual network,combined with the more popular depth residual learning method.In the residual network,when using a shortcut connection,the signal can be directly transmitted from one unit to another.Based on this,we add level-wise connections upon original residual networks,to dig the optimization ability of residual networks.For different test sets,the deep multilevel residual network model achieved better super-resolution results;both in terms of subjective vision and objective evaluation indicators,and the sharpness and edge sharpness of images are significantly improved.The experimental results show that the deep multilevel residual network has faster convergence and better reconstruction quality.