next up previous
Next: Wavelet basis and Wavelet Up: Introduction to Wavelet Networks Previous: Introduction to Wavelet Networks


Direct Calculation of Weights


Wavelet functions are not necessarily orthogonal. For a given family $ {\bf\Psi }$ of wavelets it is therefore not generally possible to calculate a wavelet coefficient $ w_i$ directly by a simple projection of the wavelet $ \psi_{{\bf n}_i}$ onto the considered function. In [Daubechies, 1990,Daugman, 1988] it was therefore proposed to use eq. (2) to find the optimal coefficients $ w_i$ for each fixed wavelet. Because optimization is a slow process, we want to suggest a direct calculation for the case of a finite wavelet family. The correct coeffs. $ w_i$ are computed by projecting the dual wavelets $ \tilde{\psi}_{{\bf n}_i}$. The wavelet $ \tilde{\psi}_{{\bf n}_i}$ is the dual wavelet to the wavelet $ \psi_{{\bf n}_i}$ if

$\displaystyle \langle\psi_{{\bf n}_i},\tilde{\psi}_{{\bf n}_j}\rangle =\delta_{i,j} \; .$ (4)

With $ \tilde{{\bf\Psi}} =(\tilde{\psi}_{{\bf n}_1},\ldots,\tilde{\psi}_{{\bf n}_N})^T$, we can write

$\displaystyle \left[\left<{\bf\Psi},\tilde{{\bf\Psi}}\right>\right]={\rm 1\negthickspace I}$ (5)

and we find $ \tilde{\psi}_{{\bf n}_i}$ to be

$\displaystyle \tilde{\psi}_{{\bf n}_i} = \sum_j\left(\Psi\right)_{i,j}^{-1}\psi_{{\bf n}_j}\; ,$ (6)

where $ ({\bf\Psi})_{i,j}=\langle\psi_i,\psi_j\rangle$. Given a family $ {\bf\Psi }$ of optimized wavelets of a WN for the function $ f$, we can compute the orthogonal projection of a function $ g$ into the subspace $ <{\bf\Psi}> \subseteq {\Bbb{L}}^2({\Bbb{R}}^2)$ (see (4)), i.e.

$\displaystyle \hat{g}= \sum^N_{i=1}w_i \psi_{{\bf n}_i}\;$with$\displaystyle \;{\bf w}= \tilde{{\bf\Psi}}g\;.$ (7)

The method to compute the orthogonal projection of a function $ g$ into the subspace $ <{\bf\Psi}>$ is mathematically equivalent to using the pseudo-inverse of $ {\bf\Psi }$ directly. However, using the dual wavelets of eq. (7) will proof to be computationally more efficient: For our tracking experiment we will have to deform the entire WN affinely, which means that the pseudo-inverse has to be recomputed. The matrix $ ({\bf\Psi})_{i,j}$, on the other hand, is invariant, except for a factor, to affine deformations of the WN, and only the projections $ \langle g,
\psi_{{\bf n}_i}\rangle$ need to be recomputed.


next up previous
Next: Wavelet basis and Wavelet Up: Introduction to Wavelet Networks Previous: Introduction to Wavelet Networks
Volker Krueger
2001-05-31