Neural network design for data compression based on Kernel PCA: Rate-distortion and complexity analysis

Abstract

This work presents a study of the properties of a non-linear vector quantization (VQ) method based on Kernel Principal Component Analysis (KPCA), focused on the complexity and viability of implementing this method in image processing. The theory supporting this method is described and then the method is compared to traditional quantization methods, as scalar quantization and entropy-constrained vector quantization. The main characteristics compared are the entropy versus distortion curves, illustrating the quantizers rate-distortion performance, and the complexity associated with the quantization process, as a function of the computational cost required in digital implementation. Finally, this work introduces a complexity-constrained approach to quantizer design.