PCA-based Seeding for Improved Vector Quantization

Abstract

We propose a new method for finding initial codevectors for vector quantization. It is based on Principal Component Analysis and uses error-directed subdivision of the eigenspace in reduced dimensionality. Additionally, however, we include shape-directed split decisions based on eigenvalue ratios to improve the visual appearance. The method achieves about the same image quality as the well-known k-means++ method, while providing some global control over compression priorities.