Full-reference image quality assessment based on the analysis of distortion process

We propose a full-reference image quality assessment metric based on the analysis of distortion process. Rather than focus on particular features of the original image and the distorted image, we attempt to assess the perceptual quality by analyzing the distortion process that degrade the original image to the distorted image. We model the distortion process as a linear mapping from the neighborhoods of an original pixel to the corresponding distorted pixel. We then employ the regularized linear regression to estimate the mapping weights. It is observed that different distortion types lead to different patterns of the mapping weights. We extract four features of the mapping weights that can represent its pattern, and employ support vector regression in order to combine them together to get the final objective score. Experimental results demonstrate that our proposed metric is more accurate than existing full-reference image quality assessment methods.