Single Image Super-resolution Using Multi-task Gaussian Process Regression

In this paper, we present a simple and novel approach for solving single image super-resolution (SISR) without external data set. Based on a variant of gaussian process regression (GPR), SISR is formulated as a multi-task regression problem in which each learning task refers to estimation of the regression function for each image patch. Unlike conventional methods, which need to specify the form of the regression function or determine many parameters in the function using inefficient method, the form of regression function in our proposed is implicit defined by the kernel function and all its model parameters can be learned from training set automatically. Experimental results demonstrate that the propose method can preserve fine details and produce natural looking results with sharp edges. Compared with the state of the art algorithms, the results of our proposed is equivalent or superior to them.