Learning a blind quality evaluation engine of screen content images

Abstract

We in this paper investigate how to blindly predict the visual quality of a screen content image (SCI). With the popularity of multi-client and remote-controlling systems, SCIs and the relevant applications have been a hot research topic. In general, SCIs contain texts or graphics in cartoons, ebooks or captures of computer screens. As for blind quality assessment (QA) of natural scene images (NSIs), it has been well established since NSIs possess certain statistical properties. SCIs however do not have reliable statistic models so far and thus the associated blind QA task is hard to be addressed. Aiming at solving this problem, we first extract 13 perceptual-inspired features with the free energy based brain theory and structural degradation model. In order to avoid the overfitting and guarantee the independence of training and testing samples, we then collect 100,000 images and use their objective quality scores computed via a high-accuracy full-reference QA method for SCIs as labels, before learning a new blind quality measure from aforementioned 13 features to the objective quality score. Experimental results performed on a large-scale screen image quality assessment database (SIQAD) demonstrate that the proposed blind quality metric has a good correlation with human perception of quality, even superior to state-of-the-art full-, reduced- and no-reference QA algorithms.