Pattern Recognition Using Neighborhood Coding

Abstract

We propose a new coding algorithm for binary images based on neighborhood relations. The shape is transformed into a set of representative vectors (position invariant) by coding each pixel according to the number of neighbors in the four directions (north, east, south, west). These neighborhood vectors are transformed into a set of codes satisfying the boundary condition imposed by the size of the image in which the shape is imbedded. A label is attached to the codes to indicate a sequential order of the pixels. The combined code and label characterize an exact shape. Thus following the label ordering and performing simple comparison of the codes an exact shape match is obtained. It is interesting to note that each shape will represent a polyomino. Neighborhood image operators are developed by applying mathematical and logical operations on the code vectors. A code reduction scheme for the purpose of information reduction and generalization of the shape image is proposed. Using the digits 1 and 0 of the NIST handwritten segmented characters set, we show a preliminary application for pattern recognition.