REVERSIBLE IMAGE COMPRESSION BOUNDED BY NOISE

Reversible image compression rarely achieves compression ratios larger than about 3 : 1. An explanation of this limit is offered, which hing es upon the additive noise the sensor introduces into the image. Simpl e models of this noise allow lower bounds on the bit rate to be estima ted from sensor noise parameters rather than from ensembles of typical images. The model predicts that an 8-b single-band image subject to n oise with unit standard deviation can be compressed reversibly to no l ess than 2.0 b/pixel, equivalent to a maximum compression ratio of abo ut 4 : 1. The model has been extended to multispectral imagery. The Ai rborne Visible and Infra Red Imaging Spectrometer (AVIRIS) is used as an example, as the noise in its 224 bands is well characterized. The m odel predicts a lower bound on the bit rate for the compressed data of about 5.5 b/pixel when a single codebook is used to encode all the ba nds. A separate codebook for each band (i.e., 224 codebooks) reduces t his bound by 0.5 b/pixel to about 5.0 b/pixel, but 90% of this reducti on is provided by only four codebooks. Empirical results corroborate t hese theoretical predictions.