Research On Image Super-Resolution Algorithm Based On Deep Learning

As one of the important ways of information expression and transmission,image plays a major role in daily life.High resolution images are rich in high-frequency details,which are widely used in medical imaging,road monitoring,satellite imaging,and so on.However,affected by imaging equipment and environmental factors,the captured images can not achieve the expected resolution.Super-resolution(SR)technology is an economic and effective way to solve this problem,and has become a hot spot in the field of information processing.In recent years,with the rapid development of neural networks,image super-resolution method based on deep learning has made a significant breakthrough in reconstruction performance,but it is still a serious challenge to quickly and effectively reconstruct high-resolution images with realistic visual effects.The most existing methods introduce more convolutional layers in exchange for the improvement of reconstruction performance,which leads to a sharp increase in model parameters and limits the application of super-resolution technology to devices with lower computing power.On the other hand,although these methods perform well in standard dataset,they are too smooth to meet the actual needs when processing real scene images.In order to overcome the above problems,this paper studies the image super-resolution problem based on the convolutional network structure and real degeneration,and proposes four effective methods to improve reconstruction efficiency and visual effects,which have certain reference significance for lightweight model designing and real scenes processing.It also has some research value for the practical application of super-resolution technology.The main contributions of this paper are as follows:1.A multi-scale channel network based on filter pruning for image super-resolution is proposed.The algorithm constructs a multi-scale channel block,which includes channel separation operation and multi-scale convolution operation to obtain accurate feature information of multiple sizes.The trained model is pruned by the filter clipping algorithm,which retaining the filters that are important for model reconstruction and cutting the unimportant filters.After pruning,the model is retrained to get the lightweight model.Experiments show that compared with other super-resolution algorithms,this method is competitive in terms of model parameters and reconstruction performance.2.A channel rearrangement multi-branch network for image super-resolution is proposed.This method rearranges the feature channels by calculating the total variation of the feature maps,and divides the rearranged features into two groups.Among them,low total variation feature maps contain less details,and high total variation feature maps contain more details.In order to reduce the number of parameters while mining more details,the low total variation feature is used for deep feature extraction.In addition,at the end of the network,this method fuses shallow information with deep feature information to avoid information loss.Experimental results show that this method has the good reconstruction performance within500 K parameters.3.A domain migration model for image super-resolution is proposed.The algorithm is composed of domain migration network and super-resolution network.The domain migration network uses real kernel and bicubic kernel to make dataset for training,which aims to eliminate the difference between real image degradation and simulated bicubic degradation.The super-resolution network constructs several SR modules to perform upsampling task respectively,so that the reconstructed image can effectively utilize the output information of each module,which improving the utilization of information.Experimental results show that this method has good reconstruction effect in both synthetic and real datasets.4.An edge preserving network for image super-resolution is proposed.The algorithm designs an image enhancement module including Gaussian and Sobel operators at the end of the network to enhance the texture details of the image while removing noise.At the same time,a channel attention module based on Laplace operator is proposed,which can enhances the difference of features between channels by extracting edge information of features,and gives more weight to features with more edge information.In addition,the content loss is introduced into the loss function,and the Laplace mixture model is used to determine the parameters of the loss function.Experimental results show that this method has good reconstruction effects for real low-resolution images with arbitrary blur.