A Research On Single Image Super-Resolution Reconstruction Algorithm

In the field of computer vision,the technology of image super-resolution reconstruc-tion has drawn a lot of attention.It aims at reconstructing clear high-resolution images from blurred low-resolution images and has wide application in many fields,such as intel-ligent monitoring,aerial remote sensing,and medical image processing.Many traditional classic algorithms including learning algorithm have shown strong performance in this field.With the development of deep learning in recent years,deep convolutional neural networks based super-resolution algorithms are engaged in a remarkable process of trans-formation.But most traditional methods are unable to take full advantage of large-scale super-resolution datasets to learn prior knowledge about image degradation and recon-struction,while deep learning based works mainly focus on networks',structure,without combining with its application area.Since the task of super-resolution itself is ill-posed,there are still many issues to be addressed,the large-scale application still has a long way to go.Considering the great potential of deep learning,this thesis mainly focuses on the structure of super-resolution networks and loss functions.The main content is as follows:1.Deep convolutional super-resolution network based on simplified residual network framework.The residual mechanism is able to solve the problem of gradient explosion and gradient vanishment in the training process of deep networks,as well as increase the speed and stability of convergence.This thesis simplifies the framework to ease the network designing and the performance is completely not affected.2.Multi-aspect improvement to the residual blocks.This thesis replaces batch nor-malization with weight normalization to avoid the dependency on batch size,and also improve the convergence stability as well as speed.This thesis employs a wide activation mechanism and channel attention mechanism to the block for the performance advan-tage.This thesis also decomposite the convolution layers in the residual blocks to enlarge model's receptive field,this strategy shows noticeable improvement while the network is relatively small or testing image is large.3.The combined use of two common pixel-wise loss function-MSE and MAE.This thesis investigates the two loss function through theoretical analysis and data experiment and studies a different combination of the two loss functions according to different network sizes.4.A loss measurement based on fast Fourier transformation.This thesis finds that fine-tuning a network with fast Fourier transformation based loss function which is calcu-lated in the frequency domain helps decrease artifacts in the prediction super-resolution images and perform better perceptual quality assurance.