A non-parametric framework for no-reference image quality assessment

In most state-of-the-art non-distortion-specific no-reference image quality assessment (NDS NR-IQA) methods, the image quality is predicted by training a regression model based on examples of distorted images and their corresponding human subjective scores. However, one drawback of these approaches is the fact that they require a training phase of the regression parameters. In this paper, a non-parametric framework for NDS NR-IQA task is presented where no training is necessary. A nearest-neighbour (NN) classifier is first employed to determine the distortion class of the test image. Once the distortion class is identified, the quality assessment prediction is then performed through and-NN regression that utilizes the differential mean opinion score (DMOS) value associated with the labelled patches within the identified class. The proposed algorithm is simple but effective. Experimental results on the LIVE IQA database show that our algorithm achieves high correlation to human perceptual measures of image quality as well as provides competitive performance to previous NDS NR-IQA algorithms.