Local gradient patterns (LGP): An effective local-statistical-feature extraction scheme for no-reference image quality assessment

Abstract

Gradient features are known to be effective for full-reference (FR) image quality assessment (IQA). However, only a few metrics utilize gradient features for no-reference (NR) IQA. To investigate the potential benefits of gradient magnitude and phase for NR-IQA, we propose a novel and effective local-statistical-feature extraction metric, namely Local Gradient Patterns (LGP), for general-purpose NR-IQA. Using a Gaussian partial derivative filter, the image is first decomposed into two complementary components: gradient magnitude and phase. The local statistical features (e.g., the conditional probability distributions) are then extracted from the complementary components, using the derived local gradient magnitude and phase pattern operators. Finally, to facilitate NR-IQA, local statistical features that convey important structural information are mapped to the subjective mean opinion score of the image, using a support vector regression (SVR) procedure. We evaluated our proposed LGP metric using images from two publicly available test databases; the results confirm that the proposed LGP metric provides predictive performance that is superior to most state-of-the-art NR-IQA metrics and has an acceptable level of computational complexity.