Research On Super Resolution Reconstruction Algorithm Based On Multi-scale Residual Dense Network And Adversarial Learning

Image has the advantages of large information storage and intuitionistic,so it plays an important role in the information era.Image resolution is one of the most important evaluation index to assess the quality of an image.The image resolution indicates the information that the image contain.Usually the higher the resolution of the image,the more information it contains.It is important to improve image resolution for information storage and utilization.Improve software algorithm is the key of super-resolution reconstruction,rather than improving hardware devices.So,it reduce research and application costs.In production and life,super-resolution reconstruction has an important application prospect in video surveillance,remote sensing and medical diagnosis etc.With the research and development of deep learning,convolutional neural network has developed rapidly.Recent researches have proven that deep convolutional neural networks can significantly improve the quality of image super-resolution.According to the current research in image super-resolution field,the evaluation of reconstructed image quality can be divided into subjective evaluation and objective evaluation.The difference between the two evaluation standards is that the subjective evaluation standard represents the visual error between reconstructed image and original image,while the objective evaluation standard represents the image element error between images.Be aimed at solving these two problems,This paper proposes a multi-scale residual dense network(MSRDN)and a super-resolved network using wasserstain generative adversarial network(SRWGAN).MSRDN and SRWGAN aimed at reducing pixel difference and visual difference respectively between reconstructed images and original images.First,most current deep convolutional neural networks based models do not fully extract features and make full use of information of each convolutional layer.In this paper,we propose a novel multi-scale residual dense network(MSRDN).We introduce convolutional kernels of different size to get more features information.This network makes full use of local and global features according to feature fusion and residual learning.Furthermore,we introduce multi-scale(up-scaling factor)model which take advantage of inter-scale correlation and reconstruct high-resolution images of different up-sample factors in a single model.This will reduce much training time and cost.Secondly,considering to improve visual performance of reconstructed images,we improve GAN according to introducing multi-scale residual dense block(MSRDB)and global feature fusion.In addition,we introduce Wasserstein GAN(WGAN)which using Wasserstein distance to represent data distribution difference to make our network easier to train.It can solve the problem of unstable training and provide accurate training processing indicators.Finally,the network is applied to remote sensing area.We Design a super resolution reconstruction prototype system for remote sensing images and reconstruct high quality remote sensing images.This also prove the application prospect of our network.