Line vector quantization using noninteger subsampled wavelet pyramids and its application in medical imaging

Abstract

We have developed two novel techniques that can improve quality and speed for wavelet based compression algorithms without major modification of the latter. We show that replacing traditional block-shaped vectors with line vectors of the same dimension (defined along a row or column of an image or its transform), significantly reduces the distortion in the reconstructed image, while accelerating coding and decoding times. To improve the performance of the clustering algorithm, we introduce a non-integer-sub-sampled wavelet pyramid. This new type of wavelet decomposition possesses certain shift-invariant properties not found in classical wavelet pyramid structures. Unlike frames and other types of mapping that introduce data redundancy into the transform in order to induce shift-invariance, our new pyramid does not introduce any data redundancy. A fast method for implementing this new pyramid is introduced. It is shown that the resulting zerotree structure is both sparser, and more efficiently coded due to the non-integer sub-sampling process. Experimental data is provided, demonstrating the performance of our proposed architecture employing line vectors. Our data also indicates that replacing the classical pyramid with this new pyramid can significantly improve performance for a wide range of quantizer designs.