Research On Single Image Super-resolution Reconstruction Based On Convolutional Neural Network

Low-cost mobile imaging devices such as smartphone cameras and portable digital cameras often capture low-resolution images,which can't satisfy the needs of practical applications.In order to break through the inherent resolution limitation of imaging equipment,as a typical ill-conditioned inverse problem,image super-resolution can recover high-resolution image from low-resolution image,which is conducive to subsequent image analysis and understanding.The technology has been widely used in image processing,computer vision and other fields,and has attracted the attention of academia and industry.Over the years,benefited from the development of artificial intelligence technology based on deep learning,the field of super-resolution has undergone a revolution.However,the existing image super-resolution methods based on deep learning have many disadvantages,such as large amount of calculation,model size and performance,which make them difficult to meet the application requirements of mobile imaging equipment.To solve this problem,we propose dual-domain convolutional neural network and hierarchical residual neural network based on deep learning,in order to build fast,lightweight and high-performance image super-resolution model.Experimental results show that the methods in this paper have achieved better results,which is usually higher than the most advanced image super-resolution method.The work of this paper is manifested in three aspects as follows:(1)Firstly,the basic knowledge of image super-resolution reconstruction and several representative methods are briefly introduced.Then,based on the practical application requirements,this paper focuses on the research of fast,lightweight and high-performance image super-resolution network model.(2)This paper proposes a dual-domain convolutional neural network with dual-domain learning strategies for image super-resolution reconstruction.By designing an efficient dual-domain butterfly modules and the dual-domain convolutional neural network architecture,the data flow interaction and compensation between the spatial domain and the frequency domainare realized.Thereby,dual-domain feature complementation and performance enhancement of the convolutional neural network are achieved.The experimental results show that the dual-domain butterfly module is an efficient network module.Compared with the traditional spatial domain convolutional neural network,the performance of the dual-domain convolutional neural network has a better image super-resolution effect.(3)In order to further improve the image resolution,combined with the above dual-domain learning strategy and introduced the wide activation function and residual learning,this paper proposes the hierarchical residual neural network,which is used for image super-resolution reconstruction,so as to perform residual learning at different levels to improve the modeling ability of the network.Experimental results show that the method can make full use of the complementary information between the spatial and frequency domain,and obtain a better super-resolution reconstruction effect under the premise of reducing the number of network layers and network model parameters.