Signal analysis based on complex wavelet signs (Q504378)

In many wavelet applications in signal and image processing, it is commonly known that interesting features of a signal are located at points of low regularity. Local regularity, such as differentiability, is usually measured in terms of the modulus of the signal's wavelet coefficients. However, the sign or more generally the phase of the wavelet coefficients has proved to be as important in image reconstruction. In this paper the authors propose a new approach to signal analysis based on the phase of the complex wavelet coefficients. The analyzing wavelets are chosen to be analytic wavelets or what the authors call complex wavelets, which are wavelets with the property that their real and imaginary parts are Hilbert transforms of each other. The focus of the paper is on the quantity \(\sigma f(b)\) which the authors call the signature of the signal \(f\) at location \(b\) and which is defined as \[ \sigma f(b)=\lim_{a\rightarrow 0^+} \text{sgn} \langle f, k_{a,b}\rangle =\lim_{a\rightarrow 0^+}\frac{\langle f, k_{a,b}\rangle }{ \left| \langle f, k_{a,b}\rangle\right| }, \] where \(f\) is a real-valued signal, \(k\) is a complex wavelet, \(a>0\) denotes the scale parameter, \(b\in \mathbb R\) is the location parameter, and \[ k_{a,b}(x)=\frac{1}{\sqrt{a}}k\left( \frac{x-b}{a} \right). \] A wavelet \(k\) is called a signature wavelet if it belongs to the Schwartz space of \(C^\infty\) rapidly decreasing functions and has a one-sided, compactly supported, non-negative Fourier transform with \[ \text{supp } \hat{k}\subset [c,d].\quad 0<c<d< \infty . \] In the definition of \(\sigma f(b)\) if the limit does not exist, it is set equal to zero. Hence, the signature is either zero or a complex number of modulus 1. The paper studies properties of \(\sigma f(b)\) and its relationship to the local behavior of \(f.\) For example, it is shown that if \(f\) has a jump discontinuity at \(b\), then \(\sigma f(b) =\pm i,\) and if \(f\) has a symmetric cusp of order \(\gamma\in (0,1],\) then \(\sigma f(b)= \pm 1. \) Signatures of other types of signals, such as random signals, are also investigated. The paper is concluded with a discretization scheme and numerical experiments.