Subjective and Objective Quality Assessment of Compressed Screen Content Videos

Abstract

With the widespread of application scenarios such as remote office and cloud collaboration, Screen Content Video (SCV) and its processing which show different characteristics from Natural Scene Video (NSV) and its processing, are increasingly attracting researcher's attention. Among these processing techniques, quality evaluation plays an important role in various media processing systems. Despite extensive research on general Image Quality Assessment (IQA) and Video Quality Assessment (VQA), quality assessment of SCVs remains undeveloped. In particular, SCVs always suffer from compression degradations in all kinds of application scenarios. In this article, we first study subjective SCV quality assessment. Specifically, we first construct a Compressed Screen Content Video Quality (CSCVQ) database with 165 distorted SCVs compressed from 11 most common screen application scenarios using the H.264, HEVC and HEVC-SCC formats. Twenty subjects were recruited to participate in the subjective test on the CSCVQ database. Then we study objective SCV quality assessment and propose a SCV quality measure. We observe that localized protruding information such as curves and dots can be well captured by the local relative standard deviation which then can be used to measure the intra-frame quality. Base on this observation, we develop a MutiScale Relative Standard Deviation Similarity (MS-RSDS) model for SCV quality evaluation. In our model, the relative standard deviation similarity between the reference and distorted SCVs is measured from frame differences between two adjacent frames, which can capture the spatiotemporal distortions accurately. A multiscale strategy is also applied to strengthen the original single-scale model. Extensive experiments are performed to compare the proposed model with the most popular and state-of-the-art quality assessment models on the CSCVQ database. Experimental results show that our proposed MS-RSDS model which has relatively low computation complexity, outperforms other IQA/VQA models.