A combined pattern recognition approach: Theory and application

Recognizing patterns is a basic activity of human beings in their daily lives. In computer science, the research area of pattern recognition involves the study of systems which can emulate this human ability. Approaches to pattern recognition can be grouped into two general categories: a statistical (or decision-theoretical) approach and a structural (or syntactic) approach. These approaches can be used to solve different pattern recognition problems. If the data involved in the patterns are easily expressed by feature vectors, the statistical approach would likely be used. For those patterns which can be broken down into subpatterns or matched against a reference pattern, it might be advantageous to use the structural approach. Because both approaches have strengths and limitations, it has been advocated for a long time that they should be combined to recognize complex patterns.;Since the 1980's some combined approaches have been proposed which can be further classified into two major branches based on their recognition strategies. The first uses both statistical and structural methods, but uses them separately and so is no longer recognized as a combined approach. The second combines both statistical and structural approaches to form a unified method which includes numerical features and statistical information in the structural parsing. There are three common problems in existing combined approaches. The first is that assigning pattern occurrence probabilities is a difficult task. the second is that they are not pattern-independent in the real applications. The third problem is that only a few practical applications have actually been implemented.;This thesis presents a new combined approach which is pattern-independent and can be more generally applied to a broader range of applications. The objective of this approach is to take advantage of the strengths of both the statistical and the structural methods while avoiding the deficiencies of existing combined approaches. The method proceeds by first extracting statistics from a feature space to establish pattern primitives. Structural information is then constructed from these primitives and is utilized in decision making. A knowledge base is used as a bridge between the statistical and structural modules, making difficult calculations of occurrence probabilities unnecessary. The approach successes the statistical method's strength of handling noise and distortion in primitive construction and the structural method's strength of using subpatterns to recognize more complex patterns.;The approach is applied to both the interpretation of remote sensing imagery and the analysis of fingerprint data. Successful results in these two very different areas demonstrate its inherent practicality and flexibility.