Study On Super-resolution Reconstruction Algorithm Of Image Based On Convolutional Neural Network

Image super-resolution reconstruction targets at reconstruct a high-resolution image from low-resolution images.A low-resolution image may correspond to multiple high-resolution images,so reconstructing high-resolution images from low-resolution images is a non-well-posed problem and a notoriously ill-posed problem.Recently,Convolution Neural Network(CNN)has been achieved excellent performance beyond other machine learning method both in the field of Computer Vision(CV)and Natural Language Processing(NLP)because of its excellent feature extraction abilities and non-linear expression capabilities,also includes image super-resolution reconstruction.Image super-resoution reconstruction algorithm based on convolution neural network was first proposed in 2015,and then showed a vigorous development trend.While achieving great success,it also brings up corresponding problems.This thesis focuses on some issues faced in image super-resolution reconstruction based on convolution neural networks field,and conducts corresponding research.The main work of this thesis are as follows:(1)Proposed an algorithm named Image Super Resolution Based on Hybrid Dilated Convolution and Lap Lacian Pyramid(HDLap SRN).HDLap SRN expand the receptive field of the network more effectively by a hybrid dilated convolution block,and it will help the network obtain more context information,so that the high-frequency features of the image can be more effectively extracted.More over the high-resolution image is gradually reconstructed by the progressive upsampling structure based on the Laplacian pyramid,which can greatly reduce the reconstruction time based on the front-end sampling method,and can also alleviate the situation of the back-end sampling when the reconstruction multiple is too large(4 times)and the accuracy of network reconstruction is greatly reduced.Finally,the innovate kind of loss function called structural loss can better combine the global information and local information of the image,which can better guide the network learning direction,so as to reconstruct a high-resolution image with rich details and clear texture.(2)Proposed Multi-Level Feature Fusion Image Super-Resolution Based on Residual Learning(MFRSR)algorithm.MFRSR drives an efficient residual block based on group convolution,it will not only ease the difficulty that deep convolution neural networks faced with,such as hard to train,degradation and gradient vanish,but also reduce the network capacity and computation,which can improve the speed of reconstruction.In order to further reduce the parameters of the model,this thesis explores an effective parameter sharing mechanism,which further reduces the model's capacity and its storage consumption.In addition,MFRSR also introduces a multi-level feature fusion mechanism based on global and local feature fusion,it can enable the features extracted from each layer of the network to be more effectively propagated and combined in the network,so that the features proposed by each layer can choose the most suitable path for the reconstruction task,and then the ability of the network to extract features at different levels has been improved.In the last,experiments have been implemented and experiment result proves that MFRSR can reconstruct a high-resolution image more close to the ground-truth one.