Image quality assessment using natural scene statistics
MetadataShow full item record
Measurement of image quality is crucial for designing image processing systems that could potentially degrade visual quality. Such measurements allow developers to optimize designs to deliver maximum quality while minimizing system cost. This dissertation is about automatic algorithms for quality assessment of digital images. Traditionally, researchers have equated image quality with image fi- delity, or the closeness of a distorted image to a ‘reference’ image that is assumed to have perfect quality. This closeness is typically measured by modeling the human visual system, or by using different mathematical criteria for signal similarity. In this dissertation, I approach the problem from a novel direction. I claim that quality assessment algorithms deal only with images and videos that are meant for human consumption, and that these signals are almost exclusively images and videos of the visual environment. Image distortions make these so-called natural scenes look ‘unnatural’. I claim that this departure from ‘expected’ characteristics could be quantified for predicting visual quality. I present a novel information-theoretic approach to image quality assessment using statistical models for natural scenes. I approach the quality assessment problem as an information fidelity problem, in which the distortion process is viewed as a channel that limits the flow of information from a source of natural images to the receiver (the brain). I show that quality of a test image is strongly related to the amount of statistical information about the reference image that is present in the test image. I also explore image quality assessment in the absence of the reference, and present a novel method for blindly quantifying the quality of images compressed by wavelet based compression algorithms. I show that images are rendered unnatural by the quantization process during lossy compression, and that this unnaturalness could be quantified blindly for predicting visual quality. I test and validate the performance of the algorithms proposed in this dissertation through an extensive study in which ground truth data was obtained from many human subjects. I show that the methods presented can accurately predict visual quality, and that they outperform current state-ofthe-art methods in my simulations.