The Daubechies D4 Wavelet Transform in C++ and Java

http://www.bearcave.com/software/java/wavelets/daubechies/index.html

I recommend using the "save as" feature of your browser to save the C++ and Java source files (I'm not sure how to reliably suppress viewing and force download with all browsers).

• The Daubechies D4 algorithm as a C++ class

• Doxygen generated documentation for the C++ version of the Daubechies D4 algorithm. This documentation was generated with this Doxygen configuration file.

• The Daubechies D4 algorithm as a Java class

Daubechies D4 wavelet transform (D4 denotes four coefficients)

I have to confess up front that the comment here does not even come close to describing wavelet algorithms and the Daubechies D4 algorithm in particular. I don't think that it can be described in anything less than a journal article or perhaps a book. I even have to apologize for the notation I use to describe the algorithm, which is barely adequate. But explaining the correct notation would take a fair amount of space as well. This comment really represents some notes that I wrote up as I implemented the code. If you are unfamiliar with wavelets I suggest that you look at the bearcave.com web pages and at the wavelet literature. I have yet to see a really good reference on wavelets for the software developer. The best book I can recommend is Ripples in Mathematics by Jensen and Cour-Harbo.

All wavelet algorithms have two components, a wavelet function and a scaling function. These are sometime also referred to as high pass and low pass filters respectively.

The wavelet function is passed two or more samples and calculates a wavelet coefficient. In the case of the Haar wavelet this is

  coefi = oddi - eveni
 or
 coefi = 0.5 * (oddi - eveni)
 

depending on the version of the Haar algorithm used.

The scaling function produces a smoother version of the original data. In the case of the Haar wavelet algorithm this is an average of two adjacent elements.

The Daubechies D4 wavelet algorithm also has a wavelet and a scaling function. The coefficients for the scaling function are denoted as h_i and the wavelet coefficients are g_i.

Mathematicians like to talk about wavelets in terms of a wavelet algorithm applied to an infinite data set. In this case one step of the forward transform can be expressed as the infinite matrix of wavelet coefficients represented below multiplied by the infinite signal vector.

     ai = ...h0,h1,h2,h3, 0, 0, 0, 0, 0, 0, 0, ...   si
    ci = ...g0,g1,g2,g3, 0, 0, 0, 0, 0, 0, 0, ...   si+1
  ai+1 = ...0, 0, h0,h1,h2,h3, 0, 0, 0, 0, 0, ...   si+2
  ci+1 = ...0, 0, g0,g1,g2,g3, 0, 0, 0, 0, 0, ...   si+3
  ai+2 = ...0, 0, 0, 0, h0,h1,h2,h3, 0, 0, 0, ...   si+4
  ci+2 = ...0, 0, 0, 0, g0,g1,g2,g3, 0, 0, 0, ...   si+5
  ai+3 = ...0, 0, 0, 0, 0, 0, h0,h1,h2,h3, 0, ...   si+6
  ci+3 = ...0, 0, 0, 0, 0, 0, g0,g1,g2,g3, 0, ...   si+7
 

The dot product (inner product) of the infinite vector and a row of the matrix produces either a smoother version of the signal (a_i) or a wavelet coefficient (c_i).

In an ordered wavelet transform, the smoothed (a_i) are stored in the first half of an n element array region. The wavelet coefficients (c_i) are stored in the second half the n element region. The algorithm is recursive. The smoothed values become the input to the next step.

The transpose of the forward transform matrix above is used to calculate an inverse transform step. Here the dot product is formed from the result of the forward transform and an inverse transform matrix row.

      si = ...h2,g2,h0,g0, 0, 0, 0, 0, 0, 0, 0, ...  ai
   si+1 = ...h3,g3,h1,g1, 0, 0, 0, 0, 0, 0, 0, ...  ci
   si+2 = ...0, 0, h2,g2,h0,g0, 0, 0, 0, 0, 0, ...  ai+1
   si+3 = ...0, 0, h3,g3,h1,g1, 0, 0, 0, 0, 0, ...  ci+1
   si+4 = ...0, 0, 0, 0, h2,g2,h0,g0, 0, 0, 0, ...  ai+2
   si+5 = ...0, 0, 0, 0, h3,g3,h1,g1, 0, 0, 0, ...  ci+2
   si+6 = ...0, 0, 0, 0, 0, 0, h2,g2,h0,g0, 0, ...  ai+3
   si+7 = ...0, 0, 0, 0, 0, 0, h3,g3,h1,g1, 0, ...  ci+3
 

Using a standard dot product is grossly inefficient since most of the operands are zero. In practice the wavelet coefficient values are moved along the signal vector and a four element dot product is calculated. Expressed in terms of arrays, for the forward transform this would be:

  ai = s[i]*h0 + s[i+1]*h1 + s[i+2]*h2 + s[i+3]*h3
 ci = s[i]*g0 + s[i+1]*g1 + s[i+2]*g2 + s[i+3]*g3
 

This works fine if we have an infinite data set, since we don't have to worry about shifting the coefficients "off the end" of the signal.

I sometimes joke that I left my infinite data set in my other bear suit. The only problem with the algorithm described so far is that we don't have an infinite signal. The signal is finite. In fact not only must the signal be finite, but it must have a power of two number of elements.

If i=N-1, the i+2 and i+3 elements will be beyond the end of the array. There are a number of methods for handling the wavelet edge problem. This version of the algorithm acts like the data is periodic, where the data at the start of the signal wraps around to the end.

This algorithm uses a temporary array. A Lifting Scheme version of the Daubechies D4 algorithm does not require a temporary. The matrix discussion above is based on material from Ripples in Mathematics, by Jensen and Cour-Harbo. Any error are mine.

Author: Ian Kaplan
Use: You may use this software for any purpose as long as I cannot be held liable for the result. Please credit me with authorship if use use this source code.

This comment is formatted for the doxygen documentation generator