Abstract
In the previous work, the LoG (Laplacian of Gaussian) signal that is the earliest stage output of human visual neural
system was suggested to be useful in image quality assessment (IQA) model design. This work considered that LoG
signal carried crucial structural information of IQA in the position of its zero-crossing and proposed a Non-shift Edge
(NSE) based IQA model. In this study, we focus on another aspect of the properties of the LoG signal, i.e., LoG whitens
the power spectrum of natural images. Here our interest is that: when exposed to unnatural images, specifically distorted
images, how does the HVS whitening this type of signals? In this paper, we first investigate the whitening filter for
natural image and distorted image respectively, and then suggest that the LoG is also a whitening filter for distorted
images to some extent. Based on this fact, we deploy the LOG signal in the task of IQA model design by applying two
very simple distance metrics, i.e., the MSE (mean square error) and the correlation. The proposed models are analyzed
according to the evaluation performance on three subjective databases. The experimental results validate the usability of
the LoG signal in IQA model design and that the proposed models stay in the state-of-the-art IQA models.
