Learning cascade regression for super-resolution image quality assessment

Super-resolution (SR) image quality assessment (SRIQA) is a fundamental topic in the literature of SR domain. Most existing SR methods usually adopt full reference (FR) metrics which need the original images as reference, to evaluate the performance of different SR algorithms and the quality of resultant SR images. However, in practice the original HR images are not available for FR-SRIQA. Therefore, it is particularly meaningful to develop a no-reference (NR) SRIQA metric to assess SR algorithms. In this paper, a novel NR-SRIQA metric comprised of a cascade two-layer regression model is presented. The newly proposed method first employs three different kinds of perceptual statistical features to measure the degradation of SR images. Then a cascade two-layer regression framework with three independent AdaBoost Decision Tree Regression (DTR) models and one ridge regression model, is developed to build the mapping relationship from the obtained statistical features to the corresponding subjective quality scores in a coarse-to-fine manner. Thorough evaluation experiments on the benchmark database proposed by Ma et al. (Comput Vis Image Understand 158:1–16 2017) confirm that the proposed NR-SRIQA metric is capable of yielding more consistent perceptual quality assessment on SR images than other state-of-the-art approaches.