| In recent years,due to the rapid development of computer technology and multimedia technology,people’s pursuit of image quality has become higher and higher,and the image quality of traditional low-resolution images has been unable to meet people’s specific needs.Therefore,the research on the super-resolution reconstruction(SR)method of the image is particularly important.How to restore the high-resolution image with richer details in the low-resolution image is the research focus in the digital image processing task.difficulty.Studies have proved that the successful application of convolutional neural network in single-frame image super-resolution reconstruction,the image reconstruction effect obtained is significantly better than traditional reconstruction algorithms.However,the existing reconstruction network model is not deep,the convolution kernel receptive field is small,image feature extraction is limited,resulting in loss of image details,gradient explosion or gradient degradation problems are prone to occur,network training difficulty increases,which affects the image reconstruction quality.This article mainly does the following work and research content.Firstly,it introduces the research background and significance of super-resolution reconstruction technology,summarizes the research status of image super-resolution reconstruction at home and abroad and image degradation models,and then briefly explains the convolutional neural network technology and some classic based Image super-resolution reconstruction method of convolutional neural network.Secondly,in view of the problems and shortcomings of the existing super-resolution reconstruction algorithm based on convolutional neural network,this paper implements an image reconstruction method based on deep residual network.The model is divided into feature extraction,nonlinear mapping of residual learning,and image reconstruction.Take the LR image as input and convolve with the first layer of convolution kernel to extract the shallow feature information of the LR image;learn the nonlinear mapping between LR image and HR image by improving the network structure of the residual unit;in the image reconstruction part,Up-sampling the residual image by deconvolution operation,and output SR image.Experiments show that the reconstruction quality of this method on natural images is significantly improved,the stability of network training is enhanced,and the structural complexity is reduced.Finally,in view of the long running time of the network model and the problem of parameter redundancy,this paper makes further improvements to the improved deep residual network model,and implements a fast image reconstruction algorithm based on the deep residual network.After the feature extraction part,a compression layer(Shrinking)is added to compress the feature dimension,reduce the number of parameters,expand the feature dimension after nonlinear mapping,reduce the computational complexity of the model,and finally perform up-sampling through deconvolution to obtain the SR image.Experiments show that compared with the improved deep residual network reconstruction method,this method is almost unaffected in the image reconstruction of most data sets,reducing the network computational complexity and running time. |
