Research On Image Super Resolution Reconstruction Based On Deep Residual Network

Image super-resolution reconstruction is a kind of digital image processing technology that can improve image resolution and recover image details.It has important application value in medical image processing,remote sensing imaging,video surveillance and other fields.With the development of deep learning and big data analysis techniques,convolutional neural networks are increasingly used in digital image processing.The purpose of image super resolution reconstruction is to learn the high-resolution details in the image while convolutional neural network has the ability to learn detailed features.This paper aims at the application of convolutional neural network in image superresolution reconstruction,studies the residual module in SRResNet network and constructs two kinds of deep residual networks for image super-resolution reconstruction.The main work of the thesis is as follows:This article first reviews the related knowledge of convolutional neural networks and deep residual networks.It mainly introduced the structure of convolutional neural network,activation function,loss function,back propagation algorithm and its variants.In addition,the network structure of deep residual network and batch normalization(BN)are introduced in detail.Then,this paper analyzes and improves the residual module in the SRResNet network by removing batch normalization in the residual module to obtain an enhanced residual module,and builds an enhanced deep residual network with depth of 35 layers based on enhanced residual module.In order to further improve the speed of network training,the paper adds a sub-pixel convolutional layer to the last layer of enhanced deep residual network.Different from the previous super-resolution reconstruction network that lowresolution image after bicubic is used as network input,enhanced deep residual network combined with subpixel convolutional layer proposed in this paper directly uses the low-resolution image as a network input.Finally,the upsampling operation in the sub-pixel convolutional layer reconstructs the high-resolution output image.In order to solve the problem that it is difficult to reconstruct the details when upsampling factor is large,this paper adds the perceptual loss to the network loss function and proposes a new network structure called PLCNN.PLCNN uses the pre-trained VGG-16 network as a loss network to obtain the perceptual loss which is added to network loss function for training the model.Compared to networks that only use L1 loss as a loss function,PLCNN can learn more semantic information of true images.The paper selects 91 images proposed by Yang et al.and DIV2 K data sets as network training set,and compares the reconstruction results of the two training sets on the public test data sets Set5,Set14 and BSD100.In addition,the real-time analysis of the enhanced deep residual network combined with sub-pixel convolutional layer is performed in this paper and the network loss curve is drawn.The reconstruction results on the public test dataset verify the effectiveness of the enhanced deep residual network combined with subpixel convolutional layer in reconstructing image details.For PLCNN network,this article designed comparison experiments with bicubic,SRCNN,and L1 loss network when the upsampling factor was 4 times.Although the PSNR/SSIM was slightly lower,PLCNN obtained a more satisfactory visual results.This also shows that the poor correlation between PSNR/SSIM and human visual effects.