njit-etd2003-081 - New Jersey Institute of Technology

njit-etd2003-081 - New Jersey Institute of Technology njit-etd2003-081 - New Jersey Institute of Technology

archives.njit.edu
from archives.njit.edu More from this publisher
20.01.2015 Views

67 3.7.1 Continuous Wavelet Transform The continuous wavelet transform maps a function f (t) onto time-scale space as This operation can be expressed in a simpler inner product notation as ψab (t) represents a family of functions obtained from a single wavelet function ψ (t) and the dilation and translation parameters a and b as where a and b are continuous. The wavelet function ψ(t) is a band-pass function. It is desired that this function have a good time and frequency localization so that f (t) is decomposed into elementary building blocks which are jointly well localized in time and frequency. The wavelet function has to satisfy the "admissibility" condition that makes it an isometry (up to a constant) of I! (R) onto 1,2 (R x R) . This requirement limits the wavelet functions which must satisfy [45][46] where ψ(Ω) is the Fourier transform of the wavelet function ψ(t). The admissibility condition is directly related to the decay of the wavelet function ψ(t) which is required to have good localization. The admissibility condition for a continuous ψ(Ω) is equivalent to a zero-mean wavelet function in time:

68 This condition forces that the wavelet function is a band pass function and decays at least as fast as |t|^1-ε in time where e is some decay constant. (In practice one needs to have much faster decay of tit (t) , in order to have good time localization). The admissibility condition assures that the "resolution of the identity" holds [45]. This guarantees that any function f(t) E 1,2 (R^n) can be reconstructed from the wavelet space as where the wavelet coefficients were defined earlier in Eq. (3.63). Its mathematical proof is given in the references [45][46]. Whenever ψ(t) is a real function, the integral limits of Ch in Eq. (3.68) are changed from 0 to 00 . Resolution of the identity ensures that the continuous wavelet transform (CWT) is complete if W f (a,b) are known for all a and b. A continuous signal f(t) is represented by a pass band function ψ(t) and its dilated and translated versions. The dilation in time leads to different resolutions in frequency. ψ(t) is a prototype window and is normally referred to as the mother wavelet. The digital implementation of the CWT can be computed directly by convolving the signal with a scaled and dilated version of the mother wavelet. For a reasonably efficient implementation, an FFT may be applied to perform the convolution [46a]. In its discrete form, a = ao^j -=n 0-3where j • and n za are integers. This is referred to as the discrete wavelet transform (DWT), which is discussed in more detail in

68<br />

This condition forces that the wavelet function is a band pass function and<br />

decays at least as fast as |t|^1-ε in time where e is some decay constant. (In practice one<br />

needs to have much faster decay <strong>of</strong> tit (t) , in order to have good time localization).<br />

The admissibility condition assures that the "resolution <strong>of</strong> the identity" holds<br />

[45]. This guarantees that any function f(t) E 1,2 (R^n) can be reconstructed from the<br />

wavelet space as<br />

where the wavelet coefficients were defined earlier in Eq. (3.63). Its mathematical pro<strong>of</strong><br />

is given in the references [45][46]. Whenever ψ(t) is a real function, the integral limits<br />

<strong>of</strong> Ch in Eq. (3.68) are changed from 0 to 00 .<br />

Resolution <strong>of</strong> the identity ensures that the continuous wavelet transform (CWT)<br />

is complete if W f (a,b) are known for all a and b. A continuous signal f(t) is<br />

represented by a pass band function ψ(t) and its dilated and translated versions. The<br />

dilation in time leads to different resolutions in frequency. ψ(t) is a prototype window<br />

and is normally referred to as the mother wavelet. The digital implementation <strong>of</strong> the<br />

CWT can be computed directly by convolving the signal with a scaled and dilated<br />

version <strong>of</strong> the mother wavelet. For a reasonably efficient implementation, an FFT may<br />

be applied to perform the convolution [46a].<br />

In its discrete form, a = ao^j -=n 0-3where j • and n za<br />

are integers. This is<br />

referred to as the discrete wavelet transform (DWT), which is discussed in more detail in

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!