Blind Image Quality Measurement via Data-Driven Transform-Based Feature Enhancement

Blind image quality measurement (BIQM) attempts to measure image perceptual quality in the absence of reference ones. Based on the types of features, BIQM methods can be divided into two categories: hand-crafted features and learning-based features. In this article, we present a data-driven transform-based feature enhancement (DFE) approach for BIQM that combines the portability of hand-crafted features with the high performance of learning-based features. Specifically, we first extract the image structural features from Karhunen–Loéve transform (KLT), phase congruency (PC), and gradient magnitude (GM) coefficients, and then natural scene statistics (NSS) feature from the local normalized coefficient. Then, we use KLT as feature enhancement process to enhance the structural and NSS features, with Weibull distribution and generalized Gaussian distribution (GGD) modeling the distributions of transform coefficient in all the frequency bands as quality-aware features. Finally, support vector regression (SVR) is adopted to map the features to subjective scores. The proposed method has been compared favorably with the state-of-the-art BIQM methods on seven widely used databases on both artificial and authentic distortion types and achieves highly competitive performance.