Abstract

A reliable and flexible model of a low-level processing stage is one of the most crucial requirements in the development of an image understanding system (IUS). In this thesis, a model for the low-level processing stage based on a new scheme of prime component decomposition is proposed. This model is then used to develop a knowledge-based image understanding system that is capable of solving many image processing problems without employing complex algorithms. A scheme for the prime component decomposition that utilizes the maximum size geometrical polygons is devised. It is shown that the optimal decomposition element in the continuous metric space has a circular shape. The decomposition operator is also optimized in the discrete metric space to deal with the actual implementation of the prime component decomposition operator, yielding square decomposition elements. The derived decomposition operator is used to extract shape elements of the objects contained in input scenes and to produce their intermediate object descriptions. In the proposed approach of shape extraction, the prime component decomposition technique is used to partition the object's interior, while a modified Sobel operator is used to detect the object's edges. The typical errors of a shape extraction process such as noise sensitivity, description errors of diagonal objects and the description errors caused by a small sampling frequency are reduced using a shape equalization approach that is based on Fourier descriptors and nonlinear interpolation. In the development of an image understanding system, a hierarchical approach of constructing the intermediate object representation is used to represent the knowledge within the system. The knowledge base of the IUS is developed as a relational multidimensional tree structure that dynamically changes the relational links among its elements. The dynamical process of creating and transforming the knowledge base is controlled by a feedback with the low-level processing stage that reduces the memory requirements of the IUS. The traditional data type definitions are extended to include the base and derived data types. These extensions effectively represent and process the time-varying knowledge of the system and increase its overall efficiency. The high-level processing stage of the IUS is implemented based on the black-board architecture with a specialized control mechanism--the agenda-based control. This control mechanism reduces the number of computational steps within the high-level processing stage by employing a selective focusing mechanism. The functional behaviour of the proposed prime component decomposition scheme and the model of the image understanding system is experimented with several application examples including the isolation and identification of stationary and time-varying objects.