(A) knowledge-based image understanding system using the blackboard model

Abstract

In this thesis, we design and implement a knowledge-based image understanding system using the blackboard model. Conventional image understanding system based on hierarchical top-down or bottom-up approaches requires a large computational work in generating features and symbols not necessarily required for the interpretation of a particular scene. Further, it is difficult to refine false interpretation caused by incomplete low-level image segmentation and to apply the system to other domains without considerable modification of the system``s configuration and interpretation strategy. The proposed system first segments the image into primary regions and then generates initial hypotheses for those candidate regions that can be reliably matched to the stored object models. Each hypothesis is then verified by additional hypotheses generation and verification, further image structures extraction and segmentation refinement. By separating domain specific knowledge as to the object models from image interpretation and control modules, our system can be improved and applied to other domains by simply adding diverse type of knowledge sources and domain specific knowledge. Using the proposed scheme, we have tried to interpret images representing outdoor scenes such as road scenes and house scenes.