In this paper we present a sub-band coder for true color images that uses an
empirically derived perceptual masking model to set the allowable quantization noiselevel
not only for each sub-band but also for each pixel in a given sub-band. The input
image is converted into YIQ space and each channel is passed through a separable
Generalized Quadrature Mirror Filterbank (GQMF). This separates the image's
frequency content into into 4 equal width bands in both the horizontal and vertical
dimension, resulting in a representation consisting of 16 sub-bands for each channel.
Using this representation, a perceptual masking model is derived for each channel.
The model incorporates spatial-frequency sensitivity, contrast sensitivity, and texture
masking. Based on the image dependent information in each sub-band and the
perceptual masking model, noise-level targets are computed for each point in a subband.
These noise-level targets are used to set the quantization levels in a DPCM
quantizer. The output from the DPCM quantizer is then encoded, using an entropybased
coding scheme, in either lxi , 1x2, or 2x2 pixel parts, based on the the statistics
in each 4x4 sub-block of a particular sub-band. One set of codebooks, consisting of
100,000 entries, is used for all images. A block elimination algorithm takes
advantage of the peaky spatial energy distribution of sub-bands to avoid using bits for
quiescent parts of a given sub-band. The resultant bitrate depends on the complexity
of the input image. For the images we use, high quality output requires bitrates from
0.25 to 1 .25 bits/pixel, while nearly transparent quality requires 0.5 to 2.5 bits/pixel.