Research On Image Super-resolution Based On Echo State Network

Image super-resolution is a technology to get high resolution image using low resolution images.Due to the limitation of imaging equipment,environmental conditions and transmission channels,low resolution images are often obtained.Low resolution image can't provide rich details and color information than high resolution image.High resolution images bring convenience to computer vision and image processing tasks such as medical diagnosis,video surveillance,text recognition and so on.In order to get rid of the restrictive factors such as hardware constraints,expensive and so on to get a clearer image,the use of digital image software processing technology to improve the resolution of the image is particularly important for today's society.In this paper,we use echo state network to study image super-resolution based on learning.The main work is as follows:Firstly,a super-resolution algorithm based on echo state network is proposed.The method is divided into the training stage and the reconstruction stage.In the training phase,the training sample image is divided into block processing,then the feature information of the image block is extracted,so the image block and the image feature information are combined into the vector matrix into the echo state network,and the mapping mode between the high and low resolution image blocks is obtained by the echo state network.In the reconstruction stage,the test image is input into the network,and the high resolution image is reconstructed by using the mapping model in the training stage.Experimental results show that the reconstructed image obtained by the proposed method is much better than other image reconstruction methods.Secondly,the image super-resolution algorithm based on ridge regression echo state network is proposed in this paper,aiming at the blurred local edge of the image reconstructed by the image super-resolution algorithm based on the echo state network.The ridge regression learning algorithm solves the problem that the linear regression algorithm in the echo state network is easy to appear the instability of the reconstructed image and the blurred edge of the image.Experimental results show that the quality of reconstructed images based on ridge regression echo state network method has been further improved.Finally,we propose an image super-resolution algorithm based on L1 norm regularization for echo state networks.In this method,L1 regularization term is introduced in the echo state network to improve the stability of the network model,which is different from the ridge regression learning algorithm.In order to simplify the complexity of the network and avoid the appearance of fitting,the echo state network is improved and the speed of the network training is accelerated.Meanwhile,in order to enhance the constraint of network training,we also introduce global reconstruction constraints based on local constraints.Experimental results show that using L1 regularization learning algorithm,the image clarity is relatively high,the detail information is relatively rich,and the quality of the image is greatly improved.