Image Understanding by Hierarchical Symbolic Representation and Inexact Matching of Attributed Graphs
We study the symbolic representation of imagery information by a powerful global representation scheme in the form of Attributed Relational Graph (ARG), and propose new techniques for the extraction of such representation from spatial-domain images, and for performing the task of image understanding through the analysis of the extracted ARG representation. To achieve practical image understanding tasks, the system needs to comprehend the imagery information in a global form. Therefore, we propose a multi-layer hierarchical scheme for the extraction of global symbolic representation from spatial-domain images. The proposed scheme produces a symbolic mapping of the input data in terms of an output alphabet, whose elements are defined over global subimages. The proposed scheme uses a combination of model-driven and data-driven concepts. The model- driven principle is represented by a graph transducer, which is used to specify the alphabet at each layer in the scheme. A symbolic mapping is driven by the input data to map the input local alphabet into the output global alphabet. Through the iterative application of the symbolic transformational mapping at different levels of hierarchy, the system extracts a global representation from the image in the form of attributed relational graphs. Further processing and interpretation of the imagery information can, then, be performed on their ARG representation. We also propose an efficient approach for calculating a distance measure and finding the best inexact matching configuration between attributed relational graphs. For two ARGs, we define sequences of weighted error-transformations which when performed on one ARG (or a subgraph of it), will produce the other ARG. A distance measure between two ARGs is defined as the weight of the sequence which possesses minimum total-weight. Moreover, this minimum-total weight sequence defines the best inexact matching configuration between the two ARGs. The global minimization over the possible sequences is performed by a dynamic programming technique, the approach shows good results for ARGs of practical sizes. The proposed system possesses the capability to inference the alphabets of the ARG representation which it uses. In the inference phase, the hierarchical scheme is usually driven by the input data only, which normally consist of images of model objects. It extracts the global alphabet of the ARG representation of the models. The extracted model representation is then used in the operation phase of the system to: perform the mapping in the multi-layer scheme. We present our experimental results for utilizing the proposed system for locating objects in complex scenes.
Date of this Version
This work was supported by the Office of Naval Research Contract No. ONR-NOOO14-79-C-0574 and the Air Force Office of Scientific Research through Honeywell Systems and Research Contract No. S49620-83-C- 1034def.