Abstract
A pattern perception system, PPS, has been implemented on the 360/67 time-sharing system with the capability of analyzing, representing, comparing, and classifying complex line patterns. Patterns are input to the system as digitized arrays and are subject to tracing by several heuristic methods based on junction characteristics. The trace yields a set of component lines. The representation is a structured pattern description in terms of these components and distinguishes the topological structure of the pattern as a whole from the primitive features (straight segments, curves, angles) that constitute each line component of the pattern. The choice of a reference component about which to cluster the description is the determining factor in the structure obtained, and a simplicity criterion is advanced to identify a plausible first choice for this role. The topological structural description acts as the initial discriminator in the comparison or matching of patterns. Since the topological description is orientation, size, and feature invariant, such information need not be retrieved from the primitive feature portion of the description unless the initial comparison is positive. This form of representation facilitates the identification of transformationally related patterns and eliminates the need for normalization of patterns. The comparison of topological structures begins at the reference component (Level 1) and proceeds from level to level (component levels are determined by their relationship to the reference). The level-by-level comparison is used to extract embedded subpatterns and serves as the basis for scene analysis in PPS. This organized structural representation is referred to as the perceived pattern and is the product of the constructive aspect of perception. All further processing achieved by PPS is accomplished through the use of this representation. Resulting descriptions can be reduced by compressing multiple examples of component interrelationships to their general recursive rule of formation. Such a reduction of recursive patterns to their common generative form is a prerequisite for efficient classification, i.e., patterns formed by or containing a random number of applications of some simple interrelationship are more easily recognized as similar when the common generation rules have been isolated. This process results in two different levels of pattern description—the unreduced specific description with all of its peculiarities and the reduced description of the general class of which it is a member. It is the general descriptions that are stored in a pattern universe as a memory of known patterns. Unknown input patterns are compared parallelly with those contained in the pattern universe and are either classified as members of a pattern universe class or used as the basis for the formation of a new class. Many of the methods used in PPS were based on findings in perceptual psychology, e.g., hierarchical representation and sequential comparison. The manner in which such concepts were incorporated into PPS and their further influences on the functioning of the system will be discussed.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have