| Nabors has claimed that the boundary based image segmentation and representation system developed by him is suitable for a wide range of image interpretation applications. The segmentation system partitions each curve in the thinned edge image into smaller curve segments and then partitions each curve segment into line segments. The output of the segmentation system is stored using a novel hierarchical representation scheme that uses both graph and tree data structures. However, methods for handling boundary discontinuity and distortion, common problems due to poor image segmentation or preprocessing, have not been developed. Moreover, the claims that the hierarchical representation facilitates the computation of many geometric features, and the development of effective and efficient image interpretation methods for diverse applications have not been fully validated.; In this dissertation, the effects of the boundary discontinuities and distortions on the hierarchical representation are analyzed, and computationally efficient algorithms for bridging discontinuities and smoothing distorted boundaries are developed. In addition, algorithms for extracting several boundary features that are useful in achieving diverse image interpretation objectives are developed including algorithms for the extraction of the general shape features such as concave-up, concave-down, local minimum, local maximum, inflection points, concavities, etc., of open and closed curves. Also, algorithms for the construction of the convex hull and the minimum bounding rectangle of a given object are developed. All algorithms produce the desired output by directly processing the hierarchical data structure produced by the image segmentation and representation system. Furthermore, the claim that the representation is general enough to support several image interpretation applications is validated through the development of image interpretation systems for content based image retrieval, shape classification and license plate recognition. All three systems receive the graph/tree hierarchical representation as input and utilize the algorithms presented in this dissertation to achieve their objectives. The research presented in this dissertation brings us one step closer to the goal of segmenting and storing the segmentation result for later use by multiple users with different individual objectives even when their image interpretation objectives are not known at the time of image segmentation and representation. |
