KL expansion is recognized as one of the most effective tools of information compression.
One of first uses of KL expansion in image coding and compression was developed in [17]
Abbas and Fahmy ([20]) applied artificial neural nets, based on KL procedure in the area of digital signal processing and in the data compression for image coding systems. Neural network implementation eliminated some problems concerning the KL transform like large computational effort. Neural network was trained using only part of the given image while the rest of the image was coded with the same set of eigenvectors. The algorithm has shown very good performance regarding the generalization capability of the model when new images have been well represented by eigenvectors produced using other images. Yamashita and Ogawa [13] successfully applied a modified KL technique for data compression in the presence of noise.
Hilay and Rubinstein [19] developed a compression algorithm based on KL procedure, which is invariant under a given transformation (2-dimensional rotation). It allows one to recognize compressed rotated images.
The KL method was also used in data compression of speech (Chen, Huo [22]). The speech signal was decomposed using the Fourier-Bessel transformation with the KL transformation of the obtained coefficients. It allowed to reduce the data rate of synthesized speech down to 11 kbit/sec with mean-square error less than 3%.