Combining Rule-based and Machine Learning Approaches for Shape Recognition

Shape recognition applications are multi-step processes that include capture, preprocessing, feature extraction, recognition, and a decision process. Each step can use approaches in the continuum from rule-based to machine learning. This paper presents example approaches, the pros and cons of these rule-based and adaptive approaches, and performance differences in biometric applications. Rule-based and machine learning approaches represent the extremes in a continuum of approaches. Each step in the shape recognition process may include algorithms that fall anywhere on this continuum or fusion that combines the results of multiple algorithms. Determining an appropriate approach for each step is generally determined by the availability of ground-truthed data for training and testing, and how much of the analysis can be automated. If analysis is based on experience and a sufficient number of ground-truthed exemplars are available then a machine learning approach is indicated. Conversely, a rule-based approach may be best if the analysis uses predictions or adequate training data is not available.