I. INTRODUCTION
With the fast development of multimedia and internet techniques, image quality assessment (IQA) has been an increasing important and challenging issue in image communication engineering [1]. In general, reference image is not available in practical applications. Therefore, no reference IQA (NR-IQA) approach becomes highly desirable. So far, a number of NR-IQA methods has been developed to measure the visual quality of the degraded image caused by different types of distortion. These methods are mainly divided into two types: one is to extract features from natural images, and then build a prediction model such as support vector regression (SVR) and multivariate version of generalized Gaussian distribution (GGD) [2]–[9]; the other is deep learning (DL)-based methods [10]–[14]. Although the DL-based methods can achieve a high prediction accuracy, training a deep network is a very complicated task. It requires huge computational resources and a large number of labelled image samples. However, the size of existing image databases labelled with mean opinion score (MOS) and difference MOS (DMOS) is very small (1000 or so).