Image coding using transform vector quantization with training set synthesis

Dorin Comaniciu, Richard Grisel

Abstract: A new paradigm that combines data modeling and vector quantization in an effective coding technique is presented. We fit a statistical model to the input data and use the best fit parameters to synthesize training vector sets with statistics similar to the input. By knowing the best-fit parameters, the decoder can synthesize the same training sets, while identical codebooks are obtained at both encoder and decoder based on the same codebook generation procedure. As a result, complete codebook adaptation is achieved with a very small increase in the bit rate. The implementation of the new technique in the transform domain produced competitive results when compared to other methods relying on vector quantization and transform coding. In particular, the image Lena was coded at 0.28 bits/pixel with a peak signal-to-noise ratio of 32.51dB.