Research Seminar - December 14, 1998

Abstract

Although many diversified algorithms analysing/interpreting contour images have been proposed, they can be generally classified into two categories: algorithms directly applicable to raw-data contour images (e.g. Hough transform and its modifications) that require no image pre-processing except edge detection; and contour tracking algorithms.

The algorithms can also be classified according to the diameter of their computational schemes: local algorithms (e.g. local template matching, B-spline techniques); and global algorithms (e.g. Hough transform, Fourier descriptors).

Generally, the local algorithms with no contour tracking are the most recommended ones. A new proposal of such an algorithm will be presented.

The proposed algorithm, which is expected to detect selected contour primitives within digital contour images, makes use of limited diameter. Such templates should be general enough to cover the whole range of configurations expected within the corresponding contour primitive. For example, the configuration of a corner would be defined by two shape parameters: the angular width and the orientation. For the analog contour templates, selected analog features are defined that depend on the shape parameters, i.e. the parameters can be computed from equations incorporating the values of the features. For an input digital contour, fragments of the corresponding diameters would be considered prospective instances of the given analog templates. First, selected digital features would be computed for a fragment. The digital features correspond to the analog ones in a sense that if a digital contour converges to an analog one, the digital features calculated for the digital contour would converge to the analog features. Then, the prospective configuration of a template is determined by applying the values of digital features to the abovementioned equations. Finally, the digital fragment is matched to the obtained configuration of the analog template. The disparity determines the level of confidence that the digital fragment should be considered an instance of the template contour.

The presented variant of the algorithm is based on features created from moments of contours and the corresponding moment invariants. The theoretical introduction will discuss the problem of convergence of digital and analog features. Then the results obtained for various shape primitives will be presented.