Super-Resolution Reconstruction Based On Deep Learning

As the main medium of information communication in life,image plays an important role in many fields of life.Super resolution(SR)is to reconstruct clear high-resolution image through low-resolution image.With the rapid development of deep learning,convolutional neural network has been widely used in super-resolution reconstruction algorithm,and achieved good results,but also has many applications in real life.Therefore,this paper will focus on the research and improvement of image super-resolution reconstruction algorithm based on depth learning.At present,in the super-resolution algorithm based on deep learning,there are many problems,such as large network volume and difficult training.This paper improves a super-resolution algorithm based on residual dense network.In view of the shortcomings of its training model,such as redundant features and too large parameter amount,it uses a specific way of sparse connection to replace the previous dense connection in the network model,to a certain extent,to avoid the problem based on the multi feature redundancy,the relationship between convolution layers and modules is analyzed.The local feature fusion in the reference module is removed,which makes the network better extract high-frequency information.The experimental results show that the improved algorithm is similar to the original network algorithm in the results,but the training difficulty of the network model is significantly reduced,the reconstruction speed of the network model is accelerated,and the model parameters are greatly reduced,making the overall performance greatly improved.In view of the problems of some details fuzzy and serious artifacts in the current super-resolution algorithm based on GAN network,this paper analyzes and improves the perception loss function in SRGAN,explains the relationship between the perception loss and the loss of mean square error for the quality of the generated image,and takes the joint loss based on the low-frequency and high-frequency features in VGG as the new perception loss,combining the mean square Error loss to improve the network,the experimental results show that the proposed algorithm can achieve better objective evaluation index,and the details of the image are clearer,the perception quality has also been significantly improved.