Unsupervised Feature Learning for Blind Quality Assessment of Super-Resolved Images

Image super-resolution is a classic image processing issue that aims to create high-resolution images from low-resolution images. Although many algorithms in this field have been proposed so far, effective quality evaluationof such images remains a challenging research area. Conventional image quality assessment measures are not sufficiently consistent with subjective judgments on these images. Therefore, it is very important to provide specific quality assessment methods for image super-resolution. In this paper, we propose a no-reference quality evaluation method for super-resolved images that, by learning a dictionary on high-resolution images and representing super-resolved blocks, produce local features that can describe super-resolution degradations well. These features are pooled togetherby a suitable pyramidal approach and produce a global feature vector of the image. These vectors and subjective quality scores are ultimately used to train a regression model. Experimental results show that this method is not only simple and high speed, but also does not require large volumes of training data and is more efficient than existing methods.