Pattern identification using line-codebooks

In this paper we propose a new pattern identification method using line-codebooks which develop from points to curved-lines while moving toward the sample distributions. This approach is based on the idea that some types of sample distribution may require higher degree codebooks in vector quantization. We first define the line-codebooks as polynomials with two end-points and the error as the summation of the distance between the samples and the line-codebook, and then formulate the learning of coefficients and the ends of the line-codebook based on the steepest descent method so as to decrease the error. Better performance than that of the method based on point-codebooks is confirmed in several experiments using spiral data