Image Super-resolution Based On Gaussian Process Regression And Its Post-processing

Image super-resolution technique has become a hot topic in the field of image processing these decades and super-resolution methods based on learning has turned out to be an important approach to solve the problem of super-resolution. However, existing super-resolution methods based on example learning usually utilize the non-probability form to build models and obtain parameters by cross-validation. So, this paper builds models based on theory of Gaussian process regression, and studies the feasibility algorithm and the post-processing of super-resolution from the perspective of uncertainty analysis. The main contributions are as follows:1) We proposed image super-resolution post-processing algorithm based on combination optimization. A fitting value based on the regression function is a scalar function. So, when reconstructing an image by the super-resolution in forms of image patches, there are multiple candidate fitting values for each pixel. In this paper, the quality of image reconstruction can be improved by learning the combination optimization of the candidate values.2) We proposed image super-resolution based post-processing algorithm based on edge attributes restrictions. Traditional super-resolution methods based on Gaussian kernel regression build model according to appearance feature of image, which ignores the effect of edges on vision. Therefore, this paper aims at modeling based on the image edge attributes and optimizing the image quality further.3) We proposed image super-resolution based on sparse solution for Gaussian process regression. This algorithm solves the feasibility problem of Gaussian process regression by building the local Gaussian process regression model. Then to get more accurate solutions, the sparse algorithm implemented to optimize the hyper-parameters in the Gaussian kernel function, as well as the initial inputs for training.The experimental results show that the super-resolution and post-processing methods, which are introduced to reconstruct a high-resolution image in this paper, are comparable to the existing popular super-resolution methods both in subjective and objective visual evaluation criteria. And our approach performs better than the other three approaches in the experiments.