Linear prediction in lossless compression of hyperspectral images

A Block-Based Inter-Band Lossless Hyperspectral Image Compressor

“…The major difference is that BH uses each predictor on a spatially adjacent region of pixels which are likely to have similar characteristics, and not on a very elongated region of pixels which aren't. BH is also particularly similar to [7], differing in that BH doesn't do a predictor design for every single pixel. Finally, many video compressors are similar to BH in that they compress blocks in the current frame (band) by referencing blocks in the previous one.…”

“…Compression also works on only a single stack at a time, but at least for frequency-sequential storage orders (like in [10]'s images) reading a whole row of stacks at once is convenient. [7] could also be made to use small buffers by storing only the rows of the image that are necessary for prediction, but their A&P methods would need to be redesigned.…”

“…BH uses each predictor on a single square block, and predicts only from the corresponding block in the adjacent previous band. At least the predictive part of BH is basically a variant of [3] and [7], as Section 2 discusses.…”

“…on a single pixel [6] transmits side-information to indicate which of a small set of predictors to use for any given pixel, while [7] and [8] design their predictors anew for each pixel. Other approaches to hyperspectral image compression are vector quantization and transform coding; [5, §2] briefly surveys many of these.…”

Lossless Predictive Compression of Hyperspectral Images

“…The final prediction is obtained as a weighted sum where the weights correspond to membership in the clusters. Mielikainen and Toivanen [18] obtain the clustering by using the LBG algorithm [19] and then obtain the coefficients for a quasi-linear predictor. Rizzo et al [20] use a membership function to separate the pixels into two groups, a group for which intraband prediction is used and a group for which interband prediction is used.…”