Академический Документы
Профессиональный Документы
Культура Документы
http://www.ceremade.dauphine.fr/~peyre/teaching/wavelets/tp1.html
1 de 5
http://www.ceremade.dauphine.fr/~peyre/teaching/wavelets/tp1.html
Examples of piecewise regular signal together with its transformed coefficients. Large coefficients are located, at each scale, near the discontinuities of the signal. Moreover, coefficients become larges at coarser scales (on the left). Finer scale are on the right, and the finest scale represent the n/2 coefficients on the right side of the transform. Inverse transforms f1 = perform_wavelet_transform(fw,Jmin,-1); % should recover the same signal
http://www.ceremade.dauphine.fr/~peyre/teaching/wavelets/tp1.html
clf; subplot(3,1,1); plot(f); axis tight; title('Original signal'); subplot(3,1,2); plot(fM); axis tight; title( ['Linear approximation with ' num2str(M) ' coefficients.']); subplot(3,1,3); plot(fT); axis tight; title( ['Non-linear approximation with ' num2str(M) ' coefficients.']);
Comparison of linear and non-linear approximation. Linear wavelet approxomation is not efficient beceause of the discontinuities in the original signal. One can see oscillation near discontinuities because we did not keep enough fine scale coefficients in order to reconstruct well these step discontinuities. The non-linear approximation performs better because it adapts itself to the signal by keeping only high coefficients, that are localised near the steps. Comparaison of the approximation errors |f-fM|^2; for both methods. % linear v_lin = cumsum( fw(end:-1:1).^2 ); v_lin = v_lin(end:-1:1); % non-linear v = sort( abs(fw) ); v_nl = cumsum( v.^2 ); v_nl = v_nl(end:-1:1); % log/log plot of the errors clf; loglog( 1:length(v_lin), v_lin, 1:length(v_lin), v_nl ); legend('linear', 'non-linear'); axis([1 length(v_lin) 1e-5 max(v_lin) ]);
3 de 5
http://www.ceremade.dauphine.fr/~peyre/teaching/wavelets/tp1.html
Decreasing of the approximation errors. The theory predicts that the linear error should decrease like M^-1, whereas the non-linear approximation should decrease like M^-alpha, where alpha is the exponent of regularity of the function outside the steps. Everything happens as if the non-linear approximation allows us to "forget" the discontinuities.
4 de 5
http://www.ceremade.dauphine.fr/~peyre/teaching/wavelets/tp1.html
Linear approximation in Fourier basis performs poorly because of discontinuities. Non-linear approximation does not improve on this because Fourier basis is not localized in space. Spectrum of such a function decays like 1/frequency.
5 de 5