Efficient but Effective Perceptual Quality Model of Screen Content Image

Screen contents are generated by computer, like text, animation or graphics. Existing image quality assessment (IQA) models of screen content image(SCI) still cannot make a good compromise between complexity and effectiveness. To tackle this problem, we propose an efficient but effective IQA model for SCI, called multi-features similarity deviation(MFSD). To ensure the low complexity of the model, we adopt conventional technology path (i.e. feature extraction, similarity measurement and feature pooling) instead of deep learning method. To ensure the effectiveness of the model, we enhance the feature measurement of gradient magnitude (SCI contains abundant sharp edges, causing the structure information sensitive for human eyes), considering luminance and chrominance (to make our method complete and robust). Compared with existing SC-IQA models, MFSD can averagely achieve the best effectiveness with low computing complexity.