No-Reference Image Quality Assessment Based on Machine Learning and Outlier Entropy Samples

A growing research is focusing on approaches for assessing image quality as a result of advancements in digital imaging. Thus, there is an increasing demand for efficient no-reference image quality assessment methods, as many real-world, everyday applications lack distortion-free, i.e., pristine versions of images. This paper presents a new no-reference image quality outlier entropy perception evaluator method for the objective evaluation of real-world distorted images based on natural scene statistics and mean subtracted and contrast normalized coefficients. Distribution of the coefficients is found useful for no-reference image quality assessment, where their characteristics are investigated here. Moreover, entropies such as Shannon and approximate entropies are found suitable for quality estimation. Recent studies show perception-based approaches that demonstrate differences in correlation with subjective assessments. Similar variations are exhibited in entropy domain showing sample outliers compared to other distorted images with different distortion levels. In order to address these variations, this work presents outlier entropy perception evaluator model based on machine learning in order to describe the diversity of distortions affecting entropy and subjective scoring. Patch extraction is employed in the approach, where distortion level is estimated. The evaluatior model is found to be efficient presenting advantages using Shannon and approximate entropies and outlier detection over available perception-based image quality evaluators. The obtained results using proposed model show significant improvements in the correlation with human perceptual quality ratings.