The chapter describes an algorithm that compresses two-dimensional data arrays, which are piece-wise smooth in one direction and have oscillating events in the other direction. Seismic, hyper-spectral and fingerprints data, for example, have such a mixed structure. The transform part of the compression process is an algorithm that combines wavelet and local cosine transform (LCT). The quantization and the entropy coding parts of the compression are taken from the SPIHT codec. To efficiently apply the SPIHT codec to a mixed coefficients array, reordering of the LCT coefficients takes place. On the data arrays, which have the mixed structure, this algorithm outperforms other algorithms that are based on the 2D wavelet transforms combined with the SPIHT coding and on the JPEG 2000 compression standard. The algorithm retains fine oscillating events even at a low bitrate. Its compression capabilities are also demonstrated on multimedia images that have a fine texture. The wavelet part in the mixed transform of the presented algorithm utilizes the library of Butterworth wavelet transforms described in Chap. 12.Currently, wavelet transforms constitute a recognized tool for signal and image processing applications. In particular, they have gained proven success in data compression. A main reason for such a success is a dual time-frequency (spatialfrequency) nature of wavelet transforms. This duality allows separation of smooth areas in a signal (image) from sharp transitions and high-frequency oscillations (edges in images). As a result, a signal (image) can be represented with a sufficient accuracy by a relatively small number of the transform's coefficients.
Spatial and Spectral Meaning of Wavelet Transform CoefficientsIn this section, we summarize some known facts that highlight the duality of the wavelet transform coefficients. Multiscale wavelet transform of a signal is implemented via iterated multirate filtering. One step in the transform of a signal x = {x[k]} consists of filtering the signal by an, approximately half-band, low-pass filterh 0 and a high-pass filterh 1 . This filtering is followed by factor-2 downsampling of both filtered signals. Thus,