Adaptive DPCM methods using linear prediction are described for the lossless compression of hyperspectral (224-band) images recorded by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS). The methods have two stages-predictive decorrelation (which produces residuals) and residual encoding. Good predictors are described, whose performance closely approaches limits imposed by sensor noise. It is imperative that these predictors make use of the high spectral correlations between bands. The residuals are encoded using variable-length coding (VLC) methods, and compression is improved by using eight codebooks whose design depends on the sensor's noise characteristics. Rice (1979) coding has also been evaluated; it loses 0.02-0.05 b/pixel compression compared with better VLC methods but is much simpler and faster. Results for compressing ten AVIRIS images are reported.

An examination of the distribution of signals within the prediction loop of an interframe video codec reveals that small random fluctuations caused by the quantizer build up over a period of time. This quantization noise circulates within the prediction loop causing a gradual decrease in the signal to noise ratio of the reconstructed sequence and an increase in the required data rate. In general, both of these effects can be nullified by the inclusion of a low pass loop filter.>

An image-coding technique is described in which an adaptive quadtree scheme is used to encode motion-compensated difference images. A mechanism is determined within the encoding process that allows a predetermined level of image quality to be selected by the user at call setup time and then maintained by the encoder. The coder operates in a variable-bit-rate mode providing a video service of near-constant image quality. A multilayered version of the coding scheme is also described in which the degradation caused by cell loss during transmission can be limited in its effect and removed completely within a short period of time. The recovery from cell loss is achieved by sending blocks from the original image sequence whenever cells are known to have been deleted by the local-network access node.>

Two modifications to the constant area quantization (CAQ) video bandwidth compression scheme are described which together reduce the mean square error (MSE) in the reconstructed image by up to 70 percent. The first modification is merely a change in the CAQ algorithm requiring no additional hardware, whereas the second modification exploits line-to-line redundancies and necessitates the inclusion of one line of memory.