Non-reference assessment of sharpness in blur/noise degraded images

Image sharpness perception is not only affected by blur but also by noise. Noise effect on perceived image sharpness is a puzzling problem since image sharpness may increase, up to a certain amount of noise, on even regions when noise is added to an image. In this paper, we propose a NR perceived sharpness metric GSVD (Gradient Singular Value Decomposition), that shows to be effective in correlating with subjective quality evaluation of images affected by either blur or noise. This metric (i) requires no training on human image quality ratings, (ii) provides comparable performance with full reference (FR) peak signal to noise ratio (PSNR) and multiscale structural similarity (MSSIM), and (iii) performs better than most of the state-of-the-art NR sharpness metrics when assessing quality in blurry image sets and noisy image sets jointly.