Single Image Super-Resolution Using Deep Convolutional Neural Network

Single Image Super-Resolution(SISR)technology can ameliorate the resolution of lowresolution images by software methods and improve the efficiency of subsequent image processing.Therefore,SISR has been a popular topic in image processing.In this thesis,we will consider how to apply the deep learning method into SISR so that a SISR model using Deep Convolutional Neural Network(DCNN)can be established.Main work and innovation of thesis are stated as follows:(1)Classic image classification and SISR models using DCNN are analyzed,and the development trend of DCNN is summarized.(2)In order to further improve the reconstruction accuracy of existing SISR methods,a highperformance Deep Convolutional Network(HDCN)model is proposed to implement SISR.Considering that many traditional SISR models cannot upscale images in alternative scale factors,cascaded HDCN model is adopted to achieve flexible magnifications.The implementation of deep edge-aware filter can limit error accumulation during repetitive upscaling and highlight the texture information.(3)In order to solve the problem that the performance curve fluctuates and the process of convergence is not stable during the training,gradual training strategy is adopted to accelerate the rate of convergence and improve the performance of the model.Considering that the texture information of the reconstructed image is prone to aliasing when the magnification is large,the optimization of the loss function is adopted to reduce the occurrence of image aliasing.(4)Ultimately,these two strategies are applied to the HDCN model.By comparing the performance evaluation results of proposed model with other classical SISR models,the effectiveness of mentioned strategies and the superiority of the proposed model are demonstrated.