Gaussian Process Regression-based Single Image Super Resolution

The image super-resolution reconstruction methods play an important role in the area of image processing.It aims at reconstructing the corresponding high-resolution image from one or more low-resolution images.Higher resolution means more high-frequency information and better visual effect from an image.But in real world,the quality of the obtained image usually cannot meet the demands of the user due to the limitations by the imaging equipment.Hence the image super-resolution reconstruction methods are of great significance for many applications.The present single-frame image super-resolution methods can be classified into interpolation-based,reconstruction-based and learning-based methods.In the thesis,the learning-based methods,especially the Gaussian-processregression-based methods,are studied.Normally,the existing Gaussian-process-regression-based methods divide the training set into training subsets with the simple Euclidean distance,hence the data in the training subsets don't confirm statistical rules.Meanwhile single training set is not generalized.These problems lead to unsatisfactory reconstruction result.To solve these problems,an optimized training set and Gaussian process regression-based method is proposed.In the training stage,G-means is used to get the more informative training subsets which obeys the Gaussian distribution,and the model is more generalized.Then the size of the training set is reduced by active-sampling.The independent Gaussian process regression model is built in each optimized training set.In the test stage,the proposed method finds the nearest training subset of each testing data to finish the reconstruction.The experimental result shows that the proposed method has better performance in quality and efficiency compared with other present methods.The traditional methods usually use single kernel to calculate the covariance matrix,but single kernel characterizes the similarity between image features by only one single distance.Hence the Gaussian process regression model can't adapt to various image features,leading to unsatisfactory SR quality.To handle this problem a spectral mixture kernel Gaussian process regression-based method is proposed.The spectral mixture kernel can characterize the similarity between image features in the form of weighted summation of various kinds of frequency components of different distance functions,and with the spectral mixture kernel the Gaussian process regression model can measure the distance between complex image features more effectively and captures more similar information,which enhances the reconstruction effect.The experimental results demonstrate that the proposed method produces more details and outperforms other traditional methods.