Jensen-Shannon divergence for visual quality assessment

This paper focuses on some theoretical properties of the Jensen-Shannon divergence (JSD) that well match human visual system (HVS) features. In particular, it is firstly shown that JSD between the probability density function (pdf) of the reference (original) image and the test (distorted) one can be reformulated as a mathematical expansion, independently of the image subject and the distortion kind. Such expansion contains a component of the well-known structural similarity measure (SSIM), some powers of the Weber's law, and an additional component tied to the skewness of the involved pdfs. This theoretical link with some HVS-based quantities is further stressed by some experimental results showing that JSD has a trend similar to other popular quality assessment measures like SSIM, VIF, and MSE on frequent degradation kinds like JPEG, Additive Gaussian Noise, Gaussian Blur, and JPEG2K. Experimental tests on several images from LIVE database show that the proposed approach can be a valid theoretical and objective support for modeling HVS behavior.