Scaling and Wavelet Filter


Y = qmf(X,P)
Y = qmf(X)
Y = qmf(X,0)


Y = qmf(X,P) changes the signs of the even index elements of the reversed vector filter coefficients X if P is 0. If P is 1, the signs of the odd index elements are reversed. Changing P changes the phase of the Fourier transform of the resulting wavelet filter by π radians.

Y = qmf(X) is equivalent to Y = qmf(X,0).

Let x be a finite energy signal. Two filters F0 and F1 are quadrature mirror filters (QMF) if, for any x,


where y0 is a decimated version of the signal x filtered with F0 so y0 defined by x0 = F0(x) and y0(n) = x0(2n), and similarly, y1 is defined by x1 = F1(x) and y1(n) = x1(2n). This property ensures a perfect reconstruction of the associated two-channel filter banks scheme (see Strang-Nguyen p. 103).

For example, if F0 is a Daubechies scaling filter with norm equal to 1 and F1 = qmf(F0), then the transfer functions F0(z) and F1(z) of the filters F0 and F1 satisfy the condition (see the example for db10):



collapse all

This example shows how to create a quadrature mirror filter associated with the db10 wavelet.

Compute the scaling filter associated with the db10 wavelet.

sF = dbwavf('db10');

dbwavf normalizes the filter coefficients so that the norm is equal to 1/2. Normalize the coefficients so that the filter has norm equal to 1.

G = sqrt(2)*sF;

Obtain the wavelet filter coefficients by using qmf. Plot the filters.

H = qmf(G);
title('Scaling (Lowpass) Filter G')
grid on
title('Wavelet (Highpass) Filter H')
grid on

Set the DWT extension mode to Periodization. Generate a random signal of length 64. Perform a single-level wavelet decomposition of the signal using G and H.

origmode = dwtmode('status','nodisplay');
n = 64;
rng 'default'
sig = randn(1,n);
[a,d] = dwt(sig,G,H);

The lengths of the approximation and detail coefficients are both 32. Confirm that the filters preserve energy.

[sum(sig.^2) sum(a.^2)+sum(d.^2)]
ans = 1×2

   92.6872   92.6872

Compute the frequency responses of G and H. Zeropad the filters when taking the Fourier transform.

n = 128;
F = 0:1/n:1-1/n;
Gdft = fft(G,n);
Hdft = fft(H,n);

Plot the magnitude of each frequency response.

hold on
grid on
title('Frequency Responses')
xlabel('Normalized Frequency')
legend('Lowpass Filter','Highpass Filter','Location','east')

Confirm the sum of the squared magnitudes of the frequency responses of G and H at each frequency is equal to 2.

sumMagnitudes = abs(Gdft).^2+abs(Hdft).^2;
[min(sumMagnitudes) max(sumMagnitudes)]
ans = 1×2

    2.0000    2.0000

Confirm that the filters are orthonormal.

df = [G;H];
id = df*df'
id = 2×2

    1.0000    0.0000
    0.0000    1.0000

Restore the original extension mode.


This example shows the effect of setting the phase parameter of the qmf function.

Obtain the decomposition low-pass filter associated with a Daubechies wavelet.

lowfilt = wfilters('db4');

Use the qmf function to obtain the decomposition low-pass filter for a wavelet. Then, compare the signs of the values when the qmf phase parameter is set to 0 or 1. The reversed signs indicates a phase shift of π radians, which is the same as multiplying the DFT by eiπ.

p0 = qmf(lowfilt,0)
p0 = 1×8

    0.2304   -0.7148    0.6309    0.0280   -0.1870   -0.0308    0.0329    0.0106

p1 = qmf(lowfilt,1)
p1 = 1×8

   -0.2304    0.7148   -0.6309   -0.0280    0.1870    0.0308   -0.0329   -0.0106

Compute the magnitudes and display the difference between them. Unlike the phase, the magnitude is not affected by the sign reversals.

ans = 1×8

     0     0     0     0     0     0     0     0


Strang, G.; T. Nguyen (1996), Wavelets and Filter Banks, Wellesley-Cambridge Press.

Extended Capabilities

C/C++ Code Generation
Generate C and C++ code using MATLAB® Coder™.

Introduced before R2006a