UC San Diego

In this dissertation the subjects of entropy coding and quality assessment in the context of natural image processing and compression have been revisited. Both subjects are amongst the most fundamental concepts which have been extensively studied under the theories of source coding and signal processing. In this dissertation, it will be demonstrated how conformity to the statistical properties of natural image data, makes it possible to estimate the entropy rate of such data with high accuracy and very low complexity. A maximum likelihood parameter estimation framework is proposed which not only is enabling the design of a fast and efficient entropy rate estimator, but also unifies the legacy rate estimation methods, namely the heuristic low-data-rate methods and the analytical high-data-rate methods. The concept of entropy rate crosses the concept of image quality measure, or distortion metric (fidelity criterion), most often under the subject of lossy source coding to measure the optimality of a compression scheme. However the distortion metrics are amongst the most basic concepts for evaluation of other image processing algorithms, beyond the image compression. Underlined by numerous publications, the need for a perceptual quality metric that reflects the perception of humans on the subject of visual quality is unanimously agreed upon. The endeavor to find a suitable image quality metric has resulted in the introduction of many image quality assessment methods. The contribution of this work on the subject of image quality is a modest step forward in unifying many of the legacy methods under a "probabilistic perceptual image quality" framework. It will be shown that different methods such as contrast sensitivity, channel decomposition and structural similarity methods are different realizations of the proposed framework. This framework not only unifies the legacy methods, but also provides means for comparing different legacy methods. Furthermore, the proposed framework creates opportunities to enhance most of the legacy perceptual image quality measures. Finally the probabilistic nature of image quality in the proposed method lends itself to extending the quality metric beyond image quality assessment with full-reference image. It also covers the quality assessment when there is no access to the reference image