This thesis presents a new image segmentation approach by integrating the segmentation and interpretation process of scene analysis. In the algorithm, the knowledge from a variety of sources are used to make inferences about the interpretations of regions, and the regions are merged in accordance with their possible interpretations.
The initial segmentation algorithm used are the well-known "Split and Merge" algorithm and segmentation algorithm using minimum spanning tree. Interpretation initialization takes as its input the region description after the initial segmentation, and it then generates the possible interpretations for each region by using the model information. The deduction of region interpretations is performed using a relaxation labeling process. Deduction proceeds by eliminating the possible region interpretations that are not consistent with any possible interpretation of an adjacent region. Finally, regions with the same unique interpretation are merged.
Experimental results are shown for some real image and test image by using a set of relational constraints as a source of knowledge.