A Knowledge-based System for Automatic Image Segmentation and Object Recognition

Doctoral thesis, 1994

The present thesis describes a system for automatic off-line scene analysis of individual images and image sequences. The analysis is based on the use of a priori procedural knowledge of the task and declarative knowledge of the scene to be analyzed. The system analyzes a scene with a domain model, including (i) a scene model, (ii) programmable algorithms via a blackboard structure, and (iii) a supervisor. The generic task is to find and delineate objects in images, which involves model and data dependent segmentation, feature extraction, and object labeling. The image is represented with statistical feature vectors describing local and global image properties, organized in an octave-based resolution pyramid. The image is segmented into regions - blobs - which are labeled on the basis on their relative positions and a priori numerical constraints on the possible labels. Additional object-label constraints are set by the existence of ridges/valleys between pairs of objects. The supervisor specifies a default processing chain of discrete actions on the image, called processing states, divided into execution and evaluation states. In the evaluation states, the results of the analysis in the execution states are compared with the domain model, and if the results are consistent, the default processing chain is continued. Otherwise a control mechanism - an exception handler - is invoked which contains condition-action pairs that trigger auxiliary processing depending on a diagnosis of the error event and the previous processing history. For sequences of images, a natural continuity constraint is used to correct possible erroneously labeled or delineated images. The scene analysis has been compared with an inductive method for simultaneous labeling and segmentation with Kohonen's unsupervised classification algorithm. The two major applications have been automatic left ventricular delineation in four-chamber ultrasound images and in gamma camera images of the heart. Other biomedical applications include the extraction of renograms from kidneys, hepatograms from livers, and left ventricular detection in x-ray cardiac images.