Research On Super-resolution Image Reconstruction Algorithm Based On Feature Enhancement

In special application scenarios such as video surveillance,medical imaging,remote sensing satellites,only low resolution(LR)images can be obtained because of the limitation of hardware,noise,etc.Super Resolution(SR)reconstruction technology recovers the details of the LR images through algorithms to make up for the lack of hardware.In recent years,SR algorithms based on Convolutional Neural Networks(CNNs)achieved excellent results by using a large number of data and using CNN to build the mapping between LR images and HR images.However,SR algorithms based on CNNs still face serious challenges,such as poor detail reconstruction,large parameters and calculations,and difficulty of training,which is difficult to apply to real-time scenarios.In order to make full use of the feature information of LR images,this thesis studies from the perspective of feature enhancement to improve the performance of SR reconstruction.The Information Distillation Network(IDN)with low computational cost and excellent performance is improved,and the Residual Dense Information Distillation Network(RD-IDN)is proposed.The main contributions of this thesis are:(1)adding the local residual unit to the feature enhancement module to avoid the loss of features and reduce the difficulty of training the network.(2)the features of the enhancement module are fused by using dense skip connections,so that the prediction module can make full use of feature information of LR images.(3)removing the bicubic interpolation branch and learn the end-to-end mapping from LR images to HR images to improve the generalization of the network.Experimental results show that the proposed method is superior to other SR algorithms in terms of reconstruction performance and computational cost.In order to utilize multi-scale features and reduce the dependence of the network on the prediction module,this thesis improves the RD-IDN and build the Pyramid Residual Dense Information Distillation Network(PRD-IDN).The main improvements are:(1)learning from the pyramid structure,multi-scale features can be extracted and enhanced by using progressive upsampling.(2)useful features are enhanced and useless features are suppressed during feature fusion.(3)changing the way of image prediction and reduce the network's dependence on the prediction module.The experimental results show that the network is more capable to restore LR image details by utilizing multi-scale features.