Image processing for moving objects
Thesis DisciplineElectrical Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
This thesis reports research on aspects of medical imaging and general image processing. As regards general image processing, techniques for estimating two-dimensional object motion (translation and rotation) from digital images are the main concern. Established techniques are reviewed, and new approaches are proposed. Several extensions are reported to the phase correlation (PC) technique introduced originally by Kuglin and Hines  for estimating translation. The superior performance of PC as compared to that of conventional correlation, is shown to be related to the concept of phase dominance, a term quite recently introduced in this laboratory in the context of Fourier phase retrieval. Three modifications to PC are found to be necessary for it to be able to operate effectively on images exhibiting imperfections of the kind encountered in the real world. The ability of PC to estimate the translation of an object, when the latter also rotates, is examined and illustrated. A technique called sequential predictive phase correlation (SPPC) is introduced. It is particularly useful for estimating the translational motion of objects recorded in sequences of Images. Emphasis is placed on the importance of descriptors which permit object rotation to be estimated, at comparatively small computational expense, with the aid of angular correlation. Two descriptors, the bispectrum (equivalent to the triple autocorrelation) and the magnitude of the conventional spectrum, are examined in detail. The chief advantage of the bispectrum is that it is completely insensitive to translational motion of the object, even though it retains all of its phase information (thereby sharpening the peaks of angular correlations) in marked contrast to the spectral magnitude. The theoretical basis for motion estimation by bispectral analysis is presented. Numerical comparisons of rotation estimation by angular correlation employing the bispectrum and the conventional spectral magnitude, in situations where the level of noise and other contamination is appreciable, show that the performance of the former is significantly superior. The main concern on the medical imaging side is with two aspects of digital coronary angiography. One of them relates to compensating for limitations of x-ray/video imaging systems. Screen brightness persistence, or lag, of video cameras typically employed in angiographic systems causes images acquired by them to be effectively multiple exposures, so that images of moving objects become significantly blurred. A method of compensating for the lag by weighted subtraction of successive images is presented. This method can achieve resolutions comparable to those conventionally obtainable for images of static objects, but with an increase in the overall contamination level. Averaging with motion compensation, however, allows the original signal-to-noise ratio to be restored. A second aspect of medical imaging examined and extended in this thesis is estimating the motion of coronary arteries for: 1) enhancing images for improving clinical diagnosis, and 2) assessing physiological function, especially coronary circulation. The key to this is shown to be appropriate compensation for arterial motion. New algorithms, based on the SPPC and lag removal techniques, are introduced. They are applied to the processing of clinical coronary angiogram sequences. The results presented herein show that these algorithms can estimate coronary arterial motion to useful accuracies. It appears that these algorithms could be usefully applied to the evaluation of coronary arterial narrowing due to coronary disease.