Blind method for noise estimation using frequency domain Natural Scene features

Abstract

Noise is a commonly encountered distortion which generally affects the high frequency regions in images. It can lead to masking effect when the level of noise is high. To distinguish between natural and noise-afflicted images and to quantify the level of degradation, several statistical features have proved to be noteworthy. In this work, a method is proposed utilizing divisive normalization based on Natural Scenes Statistics (NSS) model which closely relates to human visual perception. It extracts features from spatial as well as frequency domain. Extracted features are used to drive the Machine Learning (ML) model Gaussian Process Regression (GPR) for mapping of the scores. Several methods have been proposed till date which are compared on three databases LIVE, CSIQ and TID2013. The trained model gives the highest correlation to Human Visual System for assessment of noise in natural images.