Image and audio data are examples of domains in which DPCM (Differential
Pulse Coded Modulation) is an effective method for removing much of the
linear correlation between data samples. However, most simple DPCM schemes
cannot simultaneously cater for smooth signals, noisy signals, and
discontinuities such as edges in images. To overcome this, many adaptive
DPCM schemes have been proposed, including median predictors, gradient-based
switching predictors and gradient-based blending predictors. In this paper
we generalize the idea of blending predictors, and describe a powerful
technique for creating locally adaptive compound predictors. The result is a
prediction scheme which works well over a range of data types froms smooth
to noisy, and requires very few tunable parameters. We apply this scheme to
greyscale image data, colour image data, and audio data, and compare results
with some of the current best adaptive DPCM predictors.