Single Image Super-Resolution Based On Convolutional Neural Network

In recent years,single image super-resolution(SR)technology has been widely used in various fields,such as remote sensing image,medical imaging,video monitoring,and so on.By restoring the image details from a given low-resolution(LR)image,the corresponding high-resolution(HR)image can be reconstructed.With the development of deep learning,the convolutional neural network(CNN)-based single image SR technology has obtained great success.However,in many CNN-based image SR methods,the superior performance was achieved by training very large networks,which would induce large amount of parameters and heavy computational complexity.Therefore,those large models might not be suitable for real-world applications.To solve above problems,this paper proposes tw o efficient and effective CNN models for the single image SR task.(1)An image SR method based on multi-path wide-activated residual network(MWRN)is proposed to achieve a better trade-off between the model size and the performance.Firstly,a multi-path wide-activated residual block(MWRB)is proposed as the basic building block of MWRN.In order to increase the receptive fields and well detect the multi-scale features,MWRB combines multi-path wide-activated residual learning with dilation convolution.Secondly,a fusional channel attention(FCA)module,which contains a bottleneck layer and a multi-path wide-activated residual channel attention(MWRCA),is designed to well exploit the multi-level features in MWRN.The experiments demonstrate that,compared with the state-of-the-art methods,the proposed model MWRN is able to obtain the highest peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM)with a relatively small number of parameters.(2)An image SR method based on feature-compensated information distillation network(FCIDN)is proposed to make the model faster and lighter.Firstly,this method proposes a module named feature-compensated information distillation block(FCIDB).As the basic building block of FCIDN,FCIDB can extract the hierarchical features progressively via using the feature-compensated information distillation mechanism.In addition,to enhance the representation ability of FCIDB,the channel attention(CA)mechanism is introduced to rescale the refined features.Secondly,the information fusion unit(IFU)is designed to fully take advantage of the multiple levels of information in network.The experiments demonstrate that,compared with the state-of-the-art lightweight methods,FCIDN can achieve superior performance with less parameters and computational time.