Wavelet Folding and Decorrelation Across the Scale
Baraniuk, Richard G.
Tian, Jun Feng
The discrete wavelet transform (DWT) gives a compact multiscale representation of signals and provides a hierarchical structure for signal processing. It has been assumed the DWT can fairly well decorrelate real-world signals. However a residual dependency structure still remains between wavelet coefficients. It has been observed magnitudes of wavelet coefficients are highly correlated, both across the scale and at neighboring spatial locations. In this paper we present a wavelet folding technique, which folds wavelet coefficients across the scale and removes the across-the-scale dependence to a larger extent. It produces an even more compact signal representation and the energy is more concentrated in a few large coefficients. It has a great potential in applications such as image compression.