Font Size: a A A

Efficient Single Image Super-resolution Method Based On Gaussian Process Regression

Posted on:2022-07-05Degree:MasterType:Thesis
Country:ChinaCandidate:C GuoFull Text:PDF
GTID:2530307070956159Subject:Probability theory and mathematical statistics
Abstract/Summary:PDF Full Text Request
Image super-resolution aims to recover high-resolution images from existing lowresolution images.In the past two decades,it has become the focus of many scholars’ research,and has been widely applied in medicine,remote sensing image and image transmission.As a non-parametric regression model,Gaussian process regression has been applied to image superresolution tasks by many scholars.It has the advantages of simple model,adaptive solution of hyper-parameter and strong generalization ability.In order to improve the accuracy of model prediction,it is often necessary to use the external training set for model training,so the training set contains a large number of samples.However,due to the time complexity of Gaussian process regression model training,it takes a lot of time to apply it to image super-resolution.Aiming at the problem of the computation in image super-resolution using Gaussian process regression model,this paper carried out relevant research work to reduce the time complexity of model from two perspectives of reducing the size of training set sample and modifying model algorithm.The main research contents and innovations of this paper are as follows:(1)A weighted Gaussian process regression image super-resolution method based on random sample clustering and augmentation is proposedIn order to solve the problem of the slow learning speed of Gaussian process regression under large samples,we consider the classification learning of samples.It is also inefficient to classify large sample sets directly,so we propose an efficient clustering method.Firstly,a small number of samples are randomly selected from a large sample base for classification,but there is a possibility of missing important image information.Therefore,some samples are randomly selected from the remaining samples to add and expand the existing categories,and reasonable discrimination methods are set to determine whether the samples are "belonging to a certain category" or "separate into a category".In this way,multiple data sets with smaller sample sizes can be obtained in a short time.Secondly,after classification,the Gaussian process regression model is trained on multiple small sample sets by using the method of parallelization for samples,so as to reduce the training time of the model.In the testing stage,in order to improve the prediction accuracy of the model,the models were weighted and combined.We design a reasonable weights and add the predicted values of the test samples under each model to obtain the pixel values of the final highresolution image.Numerical experiment results show that the proposed method can greatly reduce the time consumption and ensure the quality of reconstructed images.(2)An image super-resolution method based on sparse Gaussian process regression is proposedThe time complexity of the classical Gaussian process regression model training is proportional to the cubic power of the training sample size.Because the calculation of the prediction model and the solution of the optimal hyper-parameter involves the inverse calculation of a large-scale Gram matrix.Therefore,we transform the original model solution into a sparse optimization problem.Firstly,we use the cross validation method instead of the maximum likelihood estimation method as the model fitting criterion.Besides we design the corresponding constraints and sparse priors according to the pre-experiment and theoretical analysis,and transform the model and hyper-parameter solving problem into a sparse optimization problem.Then,ADMM algorithm is used to solve the sparse optimization problem,and the calculation of matrix inverse is changed to matrix vector multiplication.Thus the time complexity is reduced.On this basis,a sparse Gaussian process regression model is constructed.Finally,the sparse Gaussian process regression model is applied to the task of image superresolution combined with the fast clustering method previously proposed.A large number of numerical experiments show that our method can effectively reduce the computation time and improve the quality of image reconstruction.
Keywords/Search Tags:Image super-resolution, Gaussian process regression, random sample clustering and augmentation, weighted model, cross validation, sparse prior
PDF Full Text Request
Related items