Research On Image Super-resolution Reconstruction Algorithm Based On Deep Learning

High-resolution images need to be obtained in many life fields.However,due to the limitations of the external surrounding environments and hardware equipments,it is difficult to obtain high-resolution images to meet the requirements,so the super-resolution reconstruction technology of images has risen and developed.The reconstructed high-resolution images used traditional image super-resolution reconstruction technology have problems such as smoothness,ringing bell and low definition.With the rapid development in recent years,deep learning is used to obtain rich multi-level feature information in images.Based on deep learning,for image super-resolution reconstruction algorithms,the research contents in this thesis are as follows:(1)In order to solve the problems of smoothing and artifacts in the reconstructed image caused by the upsampling operation before low-resolution images are input to the convolutional neural network,an Adam-optimized Convolutional Neural Network(CNN)super-resolution reconstruction algorithm is proposed.This algorithm designs an improved convolutional neural network and uses Adam optimization algorithm to optimize the network.The improved convolutional neural network is used to extract the features from low-resolution images firstly,then the extracted features are non-linearly mapped to obtain feature maps.Finally,the feature maps are deconvolved to reconstruct high-resolution images.(2)Because the number of convolutional layers of the fast convolutional neural network(FSRCNN)is small and the feature information of adjacent convolutional layers of FSRCNN lacks correlation,therefore,it is difficult to extract the deep information of images,which leads to poor effect of image super-resolution reconstruction.Aiming at these problems,this thesis proposes a super-resolution reconstruction algorithm of deep residual network with multi-level skip connections.Firstly,multi-level skip connected residual blocks are designed.Based on the multi-level skip connected residual blocks,a multi-level skip connected deep residual network is constructed to solve the problem of the relevance lack of the characteristic information of adjacent convolutional layers.Then,the stochastic gradient descent method(SGD)is used to train a deep residual network connected by multi-level skips with an adjustable learning rate strategy to obtain a super-resolution reconstruction model of the network.Finally,the low resolution image is input to the deep residual network super-resolution reconstruction model connected by multi-level skips.The predicted residual feature values are obtained through the residual blocks connected by multi-level skip,and combines the residual image and the low-resolution image are combined into a high-resolution image.(3)Aiming at the problem that image detail information is lost during the training process when only one upsampling operation is performed during image super-resolution reconstruction,this thesis proposes a progressive deep residual network super-resolution reconstruction(GDSR)algorithm.The GDSR network consists of an input layer,a NX reconstruction network and an output layer.The NX reconstruction network consists of 2X reconstruction networks.In the 2X reconstruction network,the deep residual network extracts feature information,and then the extracted features are upsampled by sub-pixel convolution.The GDSR algorithm performs N2 times of feature extraction and N2 times of upsampling for low-resolution images to obtain high-resolution images.The results of experiments verify that the proposed algorithm not only obtains good visual effects in subjective visual evaluation,but also the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)values measured in objective test evaluation have been improved.