Super Resolution Reconstruction Based On Deep Learning

With the development of science and technology,people have had a higher and higher requirement about the quality of the image.But due to the limitation of imaging equipment and other factors' influence,such as: optical blur,motion blur,down-sampling and noise,people always are difficult to obtain the ideal high resolution image,therefore,how to obtain the high resolution image has become a very popular topic in the field of computer vision.Due to the method of changing the conditions of the imaging equipment to obtain the high resolution image which will cost a lot of money,so more and more researchers begin to pay attention to the image super resolution reconstruction technology,this technology need deal with the low resolution image to obtain the high resolution image.In essence the core of the super resolution is to find the mapping relationship between the low resolution image characteristics and the high resolution image characteristics.In fact,the common method is to extract the certain features of the image,through a priori knowledge to solve an optimization problem to approach the mapping relationship.But it ignore the correlation between the features with the data and the super resolution task.Based on the above analysis,we introduce the deep convolution network into the task of the image super resolution reconstruction.First of all,with the help of the convolution network extracting the adaptive characteristics for the optical image and super resolution oriented,and by using the image neighborhood structure similarity theory and the artificial neural network integration,accurately close to the mapping relationship of the low resolution images with the high resolution image.The main work are:1.A SR method is proposed based on structure similarity and deep learning.Considering the existing methods in the process of reconstruction is not fully considering the information of the similar image patchs contained.The algorithm in this paper,in the stage of the network's input image preprocessing,through to segmentate the image into some small patchs,looking for similar image patchs for each image patch,then put them together,as a input data for the convolution neural network,last,the final high resolution image will be acquired by the convolution neural network's training.This method overcome the buzu of the neighborhood similarity information of the network's input data of the existing super resolution reconstruction algorithm based on the deep learning.2.A SR method is proposed based on integration theory and deep learning.For the problem of the oneness of the input data's information when use the convolution neural network of the existing method,in this paper,our method use the theory of artificial neural network integration,for different input data's information of the low resolution images,through a number of different convolution neural networks,with multiple different prediction results of the high resolution image,finally put the multiple predicted results together as a input data of a new convolution neural network,then we will get the high resolution image through the network's training.This method make full use of the artificial neural network integration theory and improve the quality of the image super resolution reconstruction effectively.To sum up,in this paper,based on signal processing and pattern recognition for research,put the deep learning as the main technology,combined with the structural similarity of image patchs and the artificial neural network integration theory,put forward two kinds of effective image super resolution reconstruction algorithm,these algorithms have solved the deficiency of the existing methods very good,at the same time,the edge and texture information of the high resolution images reconstructed have a good progress.