Customising cost function for optimising image quality measures performances

This article proposes two reduced reference variance/covariance-based image quality metrics using a neural network approach. The main contribution is that the proposed metrics are computationally simple and do not require the entire reference images to be calculated while still giving higher performance ranges than 18 other full-reference image quality metrics available in the literature. The first metric called error-based cost function is more accurate than most of the others, while the second metric called correlation-based cost function outperforms the others in terms of correlation and monotonicity. A comparative study has been conducted over three image quality databases including the LIVE (second release), TID2008 and CSIQ.