On Learning Based Single Image Super Resolution Reconstruction

Resolution is an important factor affecting the quality of images.There are two ways to improve image resolution.The first way is by improving the physical resolution of the image acquisition devices,which is named hardware based approach;the second way is by using machine learning technology,signal processing technology etc.without changing the imaging equipment,which is named image super-resolution reconstruction technology.It is difficult for hardware based approach to break the bottleneck of manufacturing technology.Furthermore,it is always limited by the cost and the working environment.Since image super-resolution reconstruction technique can improve image resolution without changing the imaging equipment,it has been widely used in military,medical,entertainment,public security,and so on.And it has been wildly researched by scholars all over the world.Image super-resolution reconstruction technology can be divided into three categories: interpolation-based approach,reconstruction-based approach and learning-based approach.Learning-based super-resolution approach is a hot spot in the research of image super-resolution technology.This paper focuses on some key issues of learning based approach such as the improvement of reconstruction model,the optimization of reconstruction coefficients,the amelioration of patches matching,and so on.The main contributions of this dissertation are as follows:(1)A single image super-resolution approach based on a probabilistic graphical model is proposed to solve the problem that the weight of low resolution(LR)neighborhood cannot always be consistent with that of high resolution(HR)neighborhood.In this approach,neighborhoods and the corresponding weights are obtained at the same time by searching algorithm,and then they are used as priori information in the later reconstruction process,in which the HR neighborhoods and their weights are alternatively optimized by using the probabilistic graphical model.Traditional neighbor embedding based approaches cannot get neighborhoods and their weights at the same time.Furthermore,these approaches cannot get the minimum reconstruction error only by LR patches,which will result in high computational complexity and large construction error.To solve these problems,the proposed approach replaces K-Nearest Neighbor algorithm with Iterative Nearest Neighbors algorithm in the process of neighborhood searching,in which the neighborhoods are iteratively selected by continuously compensating for the reconstruction errors.The results of experiments show that the proposed approach can effectively restore the edges and texture lost in the LR images and has good robustness against image noise.(2)A single image super-resolution approach based on sparse deformable neighbor embedding is proposed to avoid the blurred and jagged distortions caused by the direct combination of the dictionary atoms.This approach combines the idea of sparse neighbor embedding and the idea of deformable patches.It improves the efficiency of reconstruction by iterative sparse neighborhood searching and improves the quality of reconstructed results by deforming the neighborhood patch.To improve the deformation accuracy,the blurring and noise caused by interpolation are eliminated by filtering the high frequency information of the training image.LR input patch and neighborhood patch are used as a set of image sequences in the optical flow field based deformation method.The HR neighborhood patch is deformed by applying the motion filed between the input patch and the LR neighborhood patch.The results of experiments show that the reconstructed results of this approach are most similar to the original HR images.(3)A new single image super-resolution approach based on the infinite mixture model is proposed to solve the problem that the finite distribution based dictionary learning method limits the ability of the dictionary to represent the image details and avoid the reduction in the efficiency and quality of reconstruction caused by the large training dataset,.The proposed approach combines the Dirichlet Process and Gaussian Process Regression to estimate the distribution of the training patches and model the relationship between the LR and HR patches.A large number of LR/HR training patches are clustered by Dirichlet Process Mixture Model,for each cluster,the relationship between LR patches and HR patches is established by Gaussian Process Regression.The proposed approach can automatically determine the number of clusters and obtain an accurate regression model for each cluster.Furthermore,the computation complexity of mapping function is determined by the number of patches in each cluster,which is lower than that determined by the whole patches.The results of experiments on various images demonstrate that the proposed method is superior to some state-of-the-art methods.The experiment results show that the single image super-resolution approach based on sparse deformable neighbor embedding can reconstruct the best HR images from noisy LR images.The single image super-resolution approach based on the infinite mixture model has the best deburring ability and can reconstruct the best HR images from blurred LR images.