Reduced Reference Quality Assessment of Screen Content Images Rooted in Primitive Based Free-Energy Theory

With the growing popularity of portable electronic devices, such as portable computer and cellular phone, a wide variety of digital screen content images (SCIs) have drastically invaded into our daily lives. Unlike natural scene images, SCIs are typically composed of graphic and textual images, with simpler shapes, and a larger frequency of thin lines, which may lead to different viewing experience. Therefore, an accurate quality metric for SCIs which could take into account its special properties is of particular interest. In this paper, we propose a novel reduced-reference method for assessing the perceptual quality of SCIs. Specifically, the principle of free energy models the perception and understanding of images as an active reasoning process, in which the brain attempts to explain the visual scene with an internal generative model. Sparse primitive cues are explored to model the human perception of the visual scene taking account of the unique properties of SCIs and the structure of primitives (atoms in the dictionary). The difference of the prediction discrepancies between the pristine and distorted images is defined as a measurement of the image quality. Experimental results show the effectiveness of the proposed metric and it performs favorably against state-of-the-arts on the benchmark screen image quality assessment database.