No-training, no-reference image quality index using perceptual features

We propose a universal no-reference (NR) image quality assessment (QA) index that does not require training on human opinion scores. The new index utilizes perceptually relevant image features extracted from the distorted image. These include the mean phase con-gruency (PC) of the image, the entropy of the phase congruencyPC image, the entropy of the distorted image, and the mean gradient magnitude of the distorted image. Image quality prediction is accomplished by using a simple functional relationship of these features. The experimental results show that the new index accords closely with human subjective judgments of diverse distorted images. 1 Introduction Objective no-reference (NR) image quality assessment (IQA) refers to the design of algorithms that seek to judge the quality of distorted images without recourse to comparison with any reference image. Recently, several successful NR QA algorithms have been proposed. A new two-step framework for NR IQA called blind image quality index (BIQI) based on natural scene statistics (NSS) models was proposed in Ref. 1, then later refined to produce the distortion identification-based image verity and integrity evaluation (DIIVINE) index. 2 The DIIVINE index produces IQA results that accord very closely with human subjective judgments of quality when tested on large IQA databases.