Complex preprocessing for pattern recognition

Abstract

The construction of pattern recognition machines may eventually depend upon the development of highly complex preprocessors. This claim is supported by a discussion of the importance of perceptual grouping. Since complex preprocessing will assess more of the basic structure of a visual scene, internal representations will have to be more descriptive in nature. Two approches to descriptive internal representation are mentioned. Two of the author's programs are reviewed. One plays the Oriental game of GO at a human level and the other can recognize digitized hand printed characters. Both programs use a geometry preserving representation of features, so that calculations involving the features can assess the original geometry of the input. In addition, the GO program calculates groups of stones and performs other types of “complex”processing. Practical and philosophical arguments are given for the use of internal representation by pattern recognition programs.