Image Super-resolution Reconstruction Method Based On Residual Learning

In recent years,due to the rapid development of information technology,people have higher and higher requirements for signal and information processing,and image processing is an important part of information processing.Generally,image information needs to save bandwidth by reducing resolution,but the processed image cannot meet people's actual needs.Therefore,image super-resolution(SR)technology is particularly important.The principle is that in order to obtain a high-resolution(HR)image,information processing is required for a low-resolution(LR)image.Convolutional neural network(CNN)is currently the most widely used method in image super-resolution reconstruction.The original models based on convolutional neural networks all achieved better image reconstruction effects by increasing the depth of the neural network.But blindly increasing the number of network layers will cause a gradient explosion,significantly increase the computational complexity,increase the difficulty of training the network,and lose image details.However,with the wide application of residual learning,image super-resolution reconstruction has not only improved the reconstruction quality,but also significantly improved the amount of calculation.In this context,this article comprehensively analyzes some advantages and disadvantages of existing image super-resolution reconstruction algorithms.In order to improve the effect of image reconstruction,reduce the amount of calculation,and reduce the difficulty of training the network,the paper has carried out related research work.Firstly,this paper uses an image super-resolution reconstruction algorithm that combines global and local residual learning(GLRL).This algorithm can better learn the details of the image.Non-linear mapping using stacked local residual block(LRB)structure can effectively overcome the problem of image degradation.Because the information between the low-resolution image and the high-resolution image processed by the neural network is highly correlated,we can use the global residual learning(GRL)method to use most of the low-resolution information forhigh-resolution image reconstruction.On the basis of the aforementioned method,this paper further adopts a faster image super-resolution reconstruction algorithm combining global and local residual learning(FGLRL).The main purpose is to further reduce the amount of computation and reduce the difficulty of network training without affecting the reconstruction effect.The principle is to send low-resolution small images directly to the neural network and train them without enlarging them.The reason is that the computational complexity of the model will increase as the size of the space increases.After extracting features,use a 1x1 size convolution kernel to perform dimensional compression,and expand the dimensions after non-linear mapping,which can further reduce the calculation amount of the model.Finally,deconvolution is used for upsampling to enlarge the picture to the required size.This paper introduces two image super-resolution reconstruction algorithms based on residual learning.Experiments show that the use of stacked local residual blocks and global residual learning can effectively improve the reconstruction effect of the network,and the use of compression and expansion dimensions and deconvolution can further improve Effectively reduce computational complexity.