Perceptual Quality Prediction on Authentically Distorted Images Using a Bag of Features Approach

Current top-performing blind perceptual image quality prediction models are\ngenerally trained on legacy databases of human quality opinion scores on\nsynthetically distorted images. Therefore they learn image features that\neffectively predict human visual quality judgments of inauthentic, and usually\nisolated (single) distortions. However, real-world images usually contain\ncomplex, composite mixtures of multiple distortions. We study the perceptually\nrelevant natural scene statistics of such authentically distorted images, in\ndifferent color spaces and transform domains. We propose a bag of feature-maps\napproach which avoids assumptions about the type of distortion(s) contained in\nan image, focusing instead on capturing consistencies, or departures therefrom,\nof the statistics of real world images. Using a large database of authentically\ndistorted images, human opinions of them, and bags of features computed on\nthem, we train a regressor to conduct image quality prediction. We demonstrate\nthe competence of the features towards improving automatic perceptual quality\nprediction by testing a learned algorithm using them on a benchmark legacy\ndatabase as well as on a newly introduced distortion-realistic resource called\nthe LIVE In the Wild Image Quality Challenge Database. We extensively evaluate\nthe perceptual quality prediction model and algorithm and show that it is able\nto achieve good quality prediction power that is better than other leading\nmodels.\n