Research On Single Image Super-resolution Methods Based On Deep Learning

Single image super-resolution(SR)refers to enlarging a low-resolution(LR)image with unclear details to reconstruct the corresponding high-resolution(HR)image.High-resolution images can provide richer information,which is of great significance in computer vision tasks such as image understanding and image analysis.However,the results of many current super-resolution algorithms cannot meet the needs of practical applications,which have the disadvantages of lacking authenticity and finer details,great computational cost and etc.Therefore,the research on image super-resolution is still of great value.This paper studies the single image super-resolution algorithm based on deep learning and get effective improvements in efficiency and accuracy.The main work is as follows:(1)We make a thorough research on the development history and research status of SR,make a detailed analysis of the theoretical basis and typical algorithms,confirm the advantages of deep learning algorithms,and introduce the related deep learning methods in theoretical level.(2)We propose an improved SR algorithm based on very deep convolutional neural networks.In view of the high computational cost caused by using the interpolated LR image as the network input,our designed model takes the LR image without preprocessing as the input directly,and uses the sub-pixel convolutional layer in the final stage to enlarge,which improve the operating efficiency.The input layer of our model uses multiple sizes of kernels to extract LR image features to enhance richness.We combines global residual and local residual to resolve the difficulty in training convergence of deep models.Experiments demonstrate that the proposed method could improve image reconstruction quality and have better detail recovery.(3)We also propose a SR algorithm based on channel attention mechanism generative adversarial network.The general SR works based on convolutional neural network treats the information of each channel equally,making the restored image lacking texture details and high frequency components.Besides,simply taking the pixel-level closest HR image as the optimization goal will reduce the authenticity of the reconstructed image and the smoothing effect is serious.In response to the above problems,we design a generative adversarial network model for image SR.The generator uses deep residual network integrating the channel attention mechanism into the residual block.The discriminator uses a relative discriminator.The loss function is based on the combination of L1 loss,perceptual loss andadversarial loss.Training is conducted in three stages to enhance stability.Experimental results confirm that proposed method could reconstruct higher authenticity,richer texture details and better visual effects images.