Shape analysis in low resolution imagery
This thesis covers topics required for recognition of objects in images which are noisy and of low resolution. A contour extraction technique is presented which can determine a subpixel contour. This is done by modeling the shape as a polygonal object and assuming the camera performs square aperture sampling. This allows the shape recognition methods to operate on low resolution data. Examples of precise area and perimeter measurement will be given to demonstrate the effectness of this technique. Shape recognition of partially correct, distorted, and broken contour is accomplished by locating important curvature regions and matching vectors representing these locations to sets of library vectors. Shape matching proceeds hierarchically from large features, to less important features, to contour boundaries. Vertex detection is often an important part of feature extraction and shape recognition. For given contour pieces which may be disconnected and non-ordered, sets of significant vertices can be extracted in addition to different algorithm-specific sets of secondary and false vertices. A simple method based on local parameters which are measured based on the expected resolution is used to extract significant and secondary vertices without producing false vertices. The locations of these vertices are accurately preserved by the detection process, even when the contours are extracted from medium resolution (50-100 square pixels) objects. The method used to extract these features is scale invariant. The vertex features are rank ordered to a significance measure and used to hypothesize matches to library shapes. The ordering used significantly reduces the number of combinations which must be tried to find a match. The matching process is hierarchical in that significant vertex matches are then followed by secondary vertex matches and finally by a tolerance band contour match. The matching algorithms which are based on shape vectors and tolerance bands can handle disconnected contour pieces even when scaling and orientation is not known
Mitchell, Purdue University.
Off-Campus Purdue Users:
To access this dissertation, please log in to our