Research On Image Super-Resolution Reconstruction Method Based On Deep Learning

Image super-resolution is a classic computer vision problem.It recovers a high-resolution image according to a given low-resolution image.It has a wide range of applications in security imaging,medical image,image generation and other fields.In recent years,deep convolution neural network has made remarkable progress in image super-resolution.Researchers have proposed a large number of CNN-based models with excellent performance.Among them,the multi-scale residual network(MSRN)is a representative image superresolution model,which extracts and merges the multi-scale and hierarchical features in the process of image super-resolution reconstruction,and obtains good super-resolution performance.However,the super-resolution method based on MSRN has the following problems:(1)using a large-scale convolution kernel,resulting in high model complexity and large calculation,which is limited in practical applications;(2)failing to fully extract the multi-scale information in the image,using the same way to further extract all the information across channels,which hinders the expression ability of the network;(3)Using L1 loss of pixel level as loss function,the generated super-resolution image lacks high-frequency detail information.In this paper,we start with MSRN model,study the above problems and propose the corresponding model for improvement,and verify the effectiveness of the model on multiple datasets.The research contents and innovations are as follows:(1)In order to solve the problems of large complexity,large amount of calculation and low efficiency of MSRN model,MSRN is improved and a multi path residual network(MPRN)model is proposed.In the model,multi-layer with small convolution kernel is used to extract multi-scale features.Compared with the large convolution kernels,small convolution kernels can reduce complexity and computation.The features across the channel are sliced to form multiple paths,which are processed differently to extract rich multi-scale and multi-level features,so as to improve the expression ability of the network.In addition,feature fusion based on feedback is introduced into the model.The high-level features in the network are used to refine the low-level features,and improve the expression ability of low-level features and the quality of network reconstruction image.Compared with MSRN,MPRN has less parameters,better performance and higher efficiency.In this paper,experiments are carried out in set5 and other common data sets as well as remote sensing data sets to verify the effectiveness of the model.(2)In order to solve the problem that the super-resolution image generated by pixel level loss lacks high-frequency details,a new gradient loss is proposed.The loss includes three parts:content loss,structure loss and gradient adversarial loss.The content loss makes the generated image and the real image as close to the pixel value as possible to improve the objective quality of the generated image;the structure loss makes the gradient of the generated image and the real image as close as possible to enhance the edge and texture of the generated image;the gradient loss adversarial makes the generated image obtain richer high-frequency detail information and improve the perception quality of the image by using the generative adversarial network.On this basis,the enhanced multi-path residual network(EMPRN)is proposed in combination with the MPRN network proposed above,and experiments are carried out on multiple datasets to verify the effectiveness of the EMPRN model.