Image quality assessment based on textural structure and normalized noise

Traditional image quality assessments are mostly based on error analysis and the errors only stem from the absolute differences of pixel values or transform coefficients between the two compared images. With consideration of Human Vision System this paper proposes a quality assessment based on textural structure and normalized noise, SNPSNR. The time-frequency property of wavelet transform is utilized to represent images' textural structure and then the structural noise is figured as the difference between wavelet transform coefficients emphasized by textural structure. The noises on each level, i.e., each channel, are weighted by HVS. Due to the energy distribution property of wavelet transform, the noise quantity difference on each transform level is quite large and is not proportional to the influence caused by them. We normalize the structural noise on different levels by normalizing the coefficients on each level. SNPSNR computation adopting the PSNR form and the result data are fitted with Differential Mean Opinion Scores (DMOS) using logistic function. SNPSNR gains better performance when compared with MSSIM, HVSNR and PSNR.