Research On Super Resolution Of Image Based On Broad Neural Network

With the development and progress of science and technology,more and more de-vices have the ability to obtain image information.Due to the limitation of working scene,economic condition or hardware condition,the image resolution obtained by many imag-ing devices is not high enough to meet the actual requirements,which brings difficulties to the following image processing tasks.The process of reconstructing a high-resolution image from a sequence of images or a single low-resolution image by image processing is called image super-resolution.At present,the image super-resolution algorithm based on deep learning has achieved high reconstruction quality.However,training a deep learning model not only takes a long time,but also takes up a lot of computing resources,which is not conducive to practical application.To solve this problem,this paper studies the image super-resolution algorithm based on the broad neural network.Firstly,an image super-resolution algorithm based on broad learning system(BLS)and sub-pixel is proposed.In this algorithm,the high resolution images in the training samples are sampled into multiple sub-pixel images.Then the nonlinear mapping between low res-olution image and sub-pixel image is established by using the broad learning system.In the reconstruction stage,the trained width network is used to obtain the sub-pixel image corresponding to the low-resolution image,and then the sub-pixel image is restored to the high-resolution image.As the broad learning system is a single layer network,the least square algorithm is used for training,so the network speed is fast and the computational complexity is low.Sub-pixel sampling further reduces the dimensionality of the training sample data and enables the least square algorithm to process a larger training set with less memory.Experimental results show that the high resolution images reconstructed by this method not only have clear details and textures,but also have satisfactory objective evalu-ation indexes.Secondly,based on the idea of the broad learning system,a Broad-Elman neural net-work is proposed.This network increases the number of nodes in the hidden layer,and gives the least-squares algorithm of ridge regression for network training.The proposed Broad-Elman network not only has stronger nonlinear mapping ability,but also overcomes the problem that the training of traditional Elman network is slow and prone to local optimiza-tion.The accuracy of Broad-Elman used in the recognition of Norb data sets is obviously higher than other algorithms.Finally,an image super-resolution algorithm based on the Broad-Elman network is pro-posed.This algorithm uses the Broad-Elman network to establish the mapping relationship between low-resolution images and high-resolution images.Since the Elman network is a feedback network,it can learn the relationship between adjacent patches.This helps to make full use of the spatial neighborhood information between patches.Experimental results show that this method can obtain high-quality reconstruction results.