Image Quality Assessment using Benford’s Law and Natural Scene Statistics

Submission note: A thesis submitted in total fulfillment of the requirements for the degree of Doctor of Philosophy to the School of Engineering and Mathematical Sciences, College of of Science, Health and Engineering, La Trobe University, Victoria, Australia.

Image quality assessment seeks to measure the quality of an image as human visual system perceives. The aim of blind image quality assessment is to evaluate the image quality without access to the reference image. Image quality assessment plays an important role in many image processing applications such as medical imaging, image transmission, image restoration and image recognition. There has been a continuous research on image quality assessment. During the past two decades, several state-of-the-art algorithms have been developed. The aim of this thesis is develop blind image quality assessment algorithms based on the natural image statistics and Benford’s law. Blind image quality assessment is the first issue studied in this thesis. The key idea is to use the Benford’s law as a model for natural undistorted image in the transform domain. The distortion applied to an image can result in the divergence from Benford’s law. This divergence and some other statistical measures are used as features to train a model for blind image quality assessment. An efficient blind image quality assessment algorithm has been developed. The second problem investigated in this thesis is image distortion classification. The goal is to develop a distortion classification algorithm that can efficiently classify the distortion presented in an image. Distortion classification plays an important role in blind image quality assessment. The basic idea is to use the generalized Benford’s law in image distortion classification. The generalized Benford’s law and two of generalized Benford’s law parameters have been used as features to train a cubic support vector machine classifier. The proposed algorithm outperformed many state-of-the-art algorithms. A validation study on two widely used assumptions on natural scene statistics is the third problem investigated in this thesis. Although these two assumptions are used widely in image quality assessment algorithms, they are only tested on a limited number of natural images. A validation study is conducted on five natural image databases with high resolution images to investigate these assumptions. The results strongly support these assumptions.