Unsupervised Blind Image Quality Assessment based on Multi-Feature Fusion

Abstract

Image quality affects the visual experience of observers. How to accurately evaluate image quality has been widely studied by researchers. Unsupervised blind image quality assessment (BIQA) requires less prior knowledge than supervised ones. Besides, there is a trade-off between accuracy and complexity in most existing BIQA methods. In this paper, we propose an unsupervised BIQA framework that aims for both high accuracy and low complexity. To represent the image structure information, we employ Phase Congruency (PC) and gradient. After that, we calculate the mean subtracted and contrast normalized (MSCN) coefficient and the Karhunen-Loéve transform (KLT) coefficient to represent the naturalness of the images. Finally, features extracted from both the pristine and the distorted images are adopted to calculate the image quality with Multivariate Gaussian (MVG) model. Experiments conducted on six IQA databases demonstrate that the proposed method achieves better performance than the state-of-the-art BIQA methods.