Single-image super-resolution based on semi-supervised learning

Abstract

Supervised learning-based methods are popular in single-image super-resolution (SR), and the underlying idea is to learn a map from input low-resolution (LR) images to target high-resolution (HR) images based on a training set. The generalization of the learned map ensures the well performance of these methods on various test images. However, the universality of these methods weakens their specificity. To enhance the performance of learning-based methods on given test images, a semi-supervised learning-based method is firstly proposed for single-image SR. In particular, test image patches are used to learn a dictionary for defining a test-data-dependent feature space. By using the learned dictionary, all LR training samples can be mapped into the test-data-dependent feature space, which makes the information contained in the training set be understood according to the given SR task. Finally, a regression function defined on the test-data-dependent feature space is learned from the refined training samples for generating SR images. The experimental results show that more details are recovered by the proposed semi-supervised method than its supervised version, which means it is a key to balance the universality and the specificity of a regression function in learning-based SR.