Sub-band coding of images using predictive vector quantization

This paper presents a form of sub-band coding of images in which the advantages of both DPCM and vector quantization are combined in a single subband coder. A 16 band decomposition of the image frequency band is employed using a tree structure of two-dimensional separable quadrature mirror filters and decimators. The 16 parallel sub-bands created in this way are regarded as a 16-dimensional vector process and coded as such using a vector equivalent of DPCM. Here each vector is predicted as a single quantity and its residual is coded using vector quantization, this vector quantizer being nested within the vector-DPCM loop. As will be shown this technique yields a good SNR performance in the bit rate region between 0.4 and 0.7 bits per pixel, both for images inside and outside a large training set.