Image processing and knowledge-based techniques for automated quality improvement of digital images

Abstract

With its multidisciplinary nature drawing upon a great variety of areas such as mathematics, computer graphics, computer vision, visual psychophysics, optics, and computer science, the theory of Image Processing needs to be made accessible to practitioners from very diverse backgrounds, from amateur photographers to specialists in communications, medicine, or biology. The presented thesis focusses on improving quality of images as an automatic process, where we regard aesthetic, physical, perceptual and cognitive approaches (which respectively result in pleasant, identical, realistic, and detailed images) as different levels of the same problem: reproduction accuracy. Main contributions include self-contained fundamental material whose value is likely to remain applicable in a rapidly evolving body of knowledge. A basic strategy followed in its preparation was to provide a seamless integration of wellestablished theoretical concepts and their implementation using state-of-the-art software tools. We divide the problem into i) low-level algorithmic routines, which resemble early-vision stages, where prior knowledge about natural images is implicitly ‘coded’ in the algorithm itself; and ii) explicit knowledge representation for highlevel image processing tasks involving algorithm composition, execution, revision and comparison of candidate solutions. First, we perform a deep theoretical research of classical as well as emerging paradigms in image quality and its two main dimensions, noise and tone reproduction, with emphasis on information theory and signal processing, but also with inspiration from perceptual sciences and computational photography, thus providing a unifying approach. We focus on edge-preserving smoothing filters as a simple, yet very powerful low-level image processing tool to deal with noise reduction and improved tone reproduction by means of extracting intrinsic components of an image. Finally, we develop, prototype and execute selected low-level image processing operators in MATLAB. Second, in order to enable end-users to accomplish complex Image Processing tasks while at the same time limiting their cognitive and skill requirements, a system is provided in which expert’s knowledge is explicitly stated in the form of rules. Developed with classical knowledge-based techniques and finally implemented in Java, the proposed system allows easy adaptation to specific tasks by exchanging knowledge bases for different areas like computer vision, remote sensing or medical image analysis. Last, but not least, we are not only interested in solving Image Processing problems, we also want to capture, understand and share the reasoning behind so that others, including non-expert users, can use and build on it.