Research On Image Super-Resolution Algorithms Based On Deep Learning

The resolution of an image reflects the amount of information it contains.Image super-resolution(SR)algorithms take low-resolution(LR)images as inputs,and reconstruct high-resolution(HR)images through digital image processing and machine learning techniques.They have been widely used in many applications.In recent years,SR algorithms based on deep learning have made great progress.However,there still exist some shortages.Most of them ignore the scale diversity in nature images.Moreover,huge amount of model parameters limit the practicality of the algorithms.To solve these problems,we conduct the study of SR based on deep learning.The main contents are as follows:(1)A multi-scale convolutional neural network for SR is proposed.The scales of image contents in SR problems are various.Reconstruction of different components may be relevant to a diversity of neighbor sizes.The multi-scale convolutional neural network for SR contains several paths corresponding to different receptive fileds,which are helpful to utilize multi-scale features while reconstructing.Through multi-scale residual learning,we can improve the efficiency and generalization ability of the model,as well as accelerate the convergence.Experiments on public datasets demonstrate that it can achieve better performance with faster speed.(2)An effective SR algorithm based on dual path architecture is proposed.Most algorithms improve performance with growing number of model parameters,which is a limitation for application.Analysing residual network(ResNet)and dense convolutional network(DenseNet)from the perspective of higher order recurrent neural network(HORNN),we find that ResNet is helpful for feature reusage thus can reduce redundancy,while it is difficult to explore new features from previous outputs.Comparatively,DenseNet is able to explore new information,but suffers from high redundancy.The proposed network combines them together,which can explore new information with low redundancy to improve efficiency.Through a series of comparative experiments,the influence of key factors of the network are explored,and the appropriate parameters are selected.Compared with state-of-the-art methods,our algorithm can achieve better performance with comparable or even less amount of parameters.Take up-scale factor 3 as an example.Our algorithm has only 691 K parameters,while the peak signal-to-noise ratio(PSNR)of it on Set5,Set14,Urban100 and B100 datasets are 34.38 dB,30.31 dB,29.07 dB and 28.03 dB respectively.The research contents take into account the quality and efficiency of reconstruction,which have great application value.