Towards a blind image quality evaluator using multi-scale second-order statistics

Abstract

Natural image statistics have proved to be effective indicators in measuring quality degradations. Most of the current statistics-based Image Quality Assessment (IQA) metrics aim at utilizing features derived from first-order models. However, second-order statistics are also of great value in image quality prediction, which are not yet fully studied. In this paper, a Blind image Quality Evaluator based on Multi-scale Second-order Statistics (BQEMSS) is proposed. The distorted image is first transformed into an opponent color space, and then quality-aware features are extracted in multiple scales from the joint distribution of adjacent sub-band coefficients in the wavelet domain and the histogram of Gaussian derivative pattern in the spatial domain respectively. To quantify the statistical regularities between sub-band coefficients, three types of image dependencies are explored, including spatially adjacent dependency, sub-band orientation dependency and sub-band scale dependency. In the final step, features are stacked to form a feature vector and a regression module is employed to map the feature vectors into quality scores. Extensive experiments on several public image quality databases demonstrate that BQEMSS is superior over the relevant state-of-the-art general-purpose blind IQA models.