Curve denoising by multiscale singularity detection and geometric shrinkage (Q643641)

Suppose \(\gamma\) is a piecewise smooth curve in the plane parametrized by arc length, and which has been sampled at points evenly spaced along the length of the curve. However, the measured values \(z(n)\) and the actual values \(\gamma_n\) may differ due to noise. The article presents a method for obtaining an approximation to the original curve. The algorithm proceeds in two steps: the first step is to determine those sample points \(z(n)\) which correspond to corners of \(\gamma\), and the second step is to find smooth (denoised) curves that fit the sampled points between successive corners. Both steps of the method make use of the quantity \(\beta(s,t)\), which is defined to be the angle between the vectors \(\gamma(s)-\gamma(s-t)\) and \(\gamma(s+t)-\gamma(s)\). The authors show that for portions of \(\gamma\) that are smooth, \(\beta(s,t)=\kappa t+o(t^2)\), where \(\kappa\) is the curvature at \(\gamma(s)\). This observation is used to construct a statistical test for determining whether a sample \(z(n)\) is a corner point or a smooth point. The denoised curves \(\hat\gamma\) that lie between corners are then found by minimizing the norm of the difference between \(\hat\gamma\) and the sampled values \(z(n)\). Here, the norm is formed from a linear combination of a weighted \(l^2\) norm and the norm \(\|\gamma\|_\alpha^2=\sum_n\sum_k\beta_\gamma(n,k)2^{-2k(\alpha-1)}\), where \(\beta_\gamma(n,k)\) is the angle between \(\gamma_n-\gamma_{n-2^{k-1}}\) and \(\gamma_{n+2^{k-1}}-\gamma_n\). It is shown that the latter is a Sobolev-type norm, and that the angles \(\beta_\gamma\) can be interpreted as coefficients of the Haar wavelet transform applied to the derivative of \(\gamma\). The article is intended for a general mathematical audience knowledgeable in numerical analysis. The reader should be warned that there are occasional typographic errors, and the exposition is somewhat unevenly paced. A short section with experimental results is included.