Weighted Gaussian Process Regression for Single Image Super-resolution Based on Randomized Sample Clustering and Augmentation

Abstract

Gaussian process regression (GPR) model is widely used in the field of image processing due to its simplicity and strong generalization ability. However, when the size of the training data set is large in some vision tasks such as image super-resolution (SR), the training of GPR model needs extensive calculation. In this paper, we propose a fast clustering method and a weighted GPR model to solve this problem. In the proposed framework, we first employ a randomized sample clustering and augmentation method to reduce the size of training data set. And then we train GPR model on each training subset by parallelization method. Finally, we propose a weighted GPR model to combine multiple GPR models, which can effectively utilize the data of each subset and get higher reconstruction quality. Extensive experimental results show that the proposed method can reduce the computational cost efficiently and performs well in image super-resolution.