Abstract

It has long been realized that the research of image quality assessment (IQA) should take into account properties of the human visual system (HVS). Earlier attempts of HVS model based IQA methods took the reverse engineering type of approach to approximately emulate each known functional components of the HVS. Given the high complexity of the human visual brain, rather than explicitly modelling the HVS from the biological level, in this paper we develop an information theoretic model of retina, which is the very first neurological component of the HVS that acts as a bridge between the raw visual input and deeper processing mechanisms in high level vision. Since the information rate of visual signals is far beyond the processing capability of the visual brain, the design principle of the retina model is ``perception as data reduction''. The retina is believed to act as a neurological denoiser that trims down the input data load while retaining vital information for visual perception. The neurological denoiser can be used in a correctional prefiltering stage for current IQA methods. The result also provides information theoretical justification for the empirical findings in recent IQA research that proper data reduction by down-sampling/low-pass filtering improves the performance of IQA metrics. By concatenating the proposed neurological denoiser with existing IQA metrics as a preprocessing step (i.e. filtering both the original and the distorted images before calculating the full-reference quality metrics), performance of many IQA metrics, e.g. PSNR and SSIM can be effectively enhanced. We show in this paper that the neurological denoiser can enhance the accuracy of straightforward algorithms e.g. PSNR, to a level on par with state-of-the-art image quality metrics on several image quality databases.