Recognition-by-parts: a computational approach to human learning and generalization of shapes

In this paper human pattern recognition is modeled in terms of how human observers learn to describe patterns in terms of their perceived parts, their unary (part) and binary (relational) attributes and the way in which such attribute states "evidence' different classes of shapes. This approach, originally developed in the area of computer vision, is concerned with algorithms which enable the learning of shape descriptions from examples and the classification of new data (generalization) efficiently and accurately. An object in such an "evidence-based' system is represented by a set of rules, where each rule provides a certain amount of evidence for each object class in the database. The accumulated class evidence over all activated rules can then be used to determine the classification probability. We have examined how well this model reflects human perception by training observers to classify compound Gabor patterns and then testing them with versions of such patterns which were segmented (gray-level transformed) versions of the original training set. If the observers were to construct rules to define each pattern class in terms of perceived parts and their relations, then it should be expected that classification performance would generalize to these new patterns from the original set. Results confirm this hypothesis and the specific feature extraction, learning and rule generation model used to predict performance.