Quantifying image distortions using SVD

This paper describes a novel, yet effective model for automatically evaluating or quantifying the visual quality of a distorted image against its reference, based on the image statistical information. Image statistical information is embedded into the proposed model via a set of singular values, extracted using SVD, each of which characterizes the dynamics of image energy. Alterations of singular values (image energy) provide an effective indicator to quantify the degradation of image quality subjected to a wide variety of distortions. The proposed metric assesses image quality locally as well as globally and provides two different measures as outputs, i.e., graphical and numerical measures. The graphical measure, referred as distortion map, describes on how the image is deviated from its reference, through the local distribution of distortion over the whole image domain. The numerical measure is a scalar value obtained from the distortion map, which quantifies the global quality of distorted image. The proposed metric has been used to quantify the quality of natural images of various distortion types from LIVE image database, and experimental results shown justify the effectiveness and reliability of the proposed metric over the exiting SVD-based metrics.