UBC Theses and Dissertations
Feature enhancement in AVHRR imagery via probabilistic relaxation labeling methods Szczechowski, Carl
A set of algorithms is described which results in the detection, enhancement, extraction,-^, and identification of features in NOAA-9 AVHRR imagery. Sea ice leads (cracks) in ice images from the Beaufort Sea / Amundsen Gulf area are modelled as "lines" in the image-processing sense. Thermal gradients on the ocean surface are modelled as edges. Emphasis is given to enhancing the output of local line / edge detectors in order to provide improved input for line / edge tracking algorithms. As a result the identification scheme operates on a segmented version of the image rather than on a pixel by pixel basis, thereby providing a less noisy classification. Line / edge enhancement is achieved using the non-linear probabilistic relaxation model of Rosenfeld et al (1976). Results from the relaxation of line detector output suggests that an expanded label (i.e. line orientation) set is preferable to the smaller set suggested by previous studies. Also, a modified form of the original non-linear model (suggested by Peleg and Rosenfeld, 1978) was found to speed up the convergence rate significantly with no degradation in enhancement. A unique set of subjectively-derived compatibility coefficients was introduced into the relaxation with encouraging results. Edge relaxation using the 3x3 template matching edge operators resulted in a relatively poor enhancement due primarily to the vagaries of the edge operator. Improved results were achieved using the "crack" edge representation. A unique relaxation model consisting of a combination of the relaxation models of Prager (1980) and Hanson et al (1980) provided a good enhancement. Classification (i.e. identification) results were excellent for the sea ice lead application. Line brightness and original line detector strength seemed sufficient for differentiating between "strong" leads and "weak" leads as well as discriminating leads from clouds. The classification results were not as good in the edge case indicating that simple characteristics such those used in the line case are not sufficient for classification purposes.
Item Citations and Data