Research On Image Super-Resolution Algorithms Based On The Dense Networks

Image super-resolution generally refers to the improvement of image resolution,such as zooming from 512 × 512 resolution image to 1024 × 1024 resolution image.In recent years,the demand for high-resolution images is increasing,and the image acquisition is limited by the imaging environment,the cost of imaging equipment(such as camera),the transmission flow bandwidth,the size of hard disk storage and so on.Therefore,it is necessary to reconstruct a high-resolution image with more details through image super-resolution technology.In addition,the technology has a wide range of applications in HDTV,weather detection,video surveillance,medical images and so on.Image super-resolution algorithms can be divided into three main categories: interpolation based,reconstruction based and learning based algorithms.At present,the mainstream method is based on learning,which benefits from the powerful reasoning ability of deep learning,and the resulting image is better than other traditional methods.In this paper,we mainly do some research on the method based on deep learning:(1)At present,the image super-resolution method based on deep learning does not make full use of two aspects of information at the same time.One is that it does not make full use of the characteristics of each convolution layer;the other is that the low-resolution input features which contain abundant low-frequency information are treated equally in each channel and these channel information is not maken full use of.Therefore,dense and spatial attention block(DSAB)is proposed in this paper.DSAB can not only fuse layer to layer features,but also treat different channel features differently through spatial attention mechanism,so as to improve the network representation ability.At the same time,in order to extract more features and reduce training parameters,all features extracted by dense spatial attention blocks are fused,and then convolution operation is used,and then the output information is adaptively controlled.The experimental results show that the effect of the proposed network model is better than the current representative methods,and the edge of the image is clearer and sharper in visual effect.(2)In view of the fact that the network is too deep to train,the features are easy to be lost in the transmission process,and the features of each build-up layer entering the next build-up layer are treated without difference,this paper proposes an enhanced information distillation compression block(EDCB)structure.EDCB consists of three parts: feature extraction,information distillation and compression,and residual learning.Firstly,the features are extracted through multiple convolution layers,and then information distillation and compression are carried out.Part of the obtained features are split,and the cut features are superimposed on the input feature blocks for superposition and convolution.The remaining features are continuously convoluted to extract features.Then,the features superimposed by the previous cutting and input are added to the information distillation.Compression can not only fuse part of the extracted features with the input features,but also extract more useful information.Finally,in order to learn the high-frequency features better,all the features extracted from the enhanced information distillation compression block are used for local and global residual learning.The experimental results show that it is superior to IDN algorithm and several representative image super-resolution algorithms in objective evaluation index(PSNR or SSIM)and visual effects.