No-Reference Perceptual Image Sharpness Index Using Normalized DCT-based Representation

• Published in: 2014 Seventh International Symposium on Computational Intelligence and Design Vol. 2; pp. 150 - 153
• Main Authors: Shuhong Jiao, Huan Qi, Weisi Lin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2014
• Subjects: discrete cosine transform; Discrete cosine transforms; Image quality; Indexes; Measurement; natural scene statistics; normalization; Probability; sharpness metric
• ISBN: 9781479970049; 1479970042
• DOI: 10.1109/ISCID.2014.50

This paper presents a no-reference (NR) image sharpness algorithm based on natural scene statistics (NSS) in discrete cosine transform (DCT) domain. It relies on the assumption that natural images possess certain statistics that will change with blur distortion. We propose a new image representation, normalized discrete cosine transform (NDCT) coefficients. Both the theoretical analysis and experimental tests have proven that the statistics of NDCT coefficients are highly correlated with the human judgments of image quality. To represent the statistics of natural images, a model is built with a small set of natural images. We define the perceptual sharpness index on normalized discrete cosine transform coefficients (NDCT-PSI) as the difference between the NSS model and the tested image. The NDCT-PSI outperforms recent relevant state-of-the-art algorithms as evaluated on a subject-rated image database. The new framework we proposed here is a simple way to facilitate some practical applications.