A comparison of lossy compression methods on still and hyperspectral images

摘要

The statistical dependency across levels (or scales) in the wavelet transform of natural images can be effectively exploited to design state-of-the-art image coders. This paper presents a comparative evolution of three such algorithms, although only one of these algorithms is detailed to a larger extend. The first method explicitly models the conditional statistics of the wavelet coefficients for bit-plane encoding. The second method performs an implicit sorting of the coefficients by sending zerotree symbols. Both methods produce an embedded code, i.e. the optimal image at any rate can be identified with the truncated bitstream of the highest possible rate, but the second algorithm does not depend as much as the first on the use of arithmetic coding. The third method is based on lattice vector quantization: statistical dependency within the wavelet transform is taken into account by conditioning the encoding of the vector norm on the value of a predictor. All three algorithms yield approximately comparable experimental results for different corpuses of images. This supports our view that the essence of high performance image compression is a careful modeling of the conditional image statistics.