Ridges in image and data analysis (Q1353378)

The concept of ridges as manifolds of critical points is a natural extension of the concept of local maximum as an isolated critical point. (A similar concept of courses generalizes local minima, but since local minima of \(f\) are local maxima of \(-f\), it is sufficient to study only the concept of ridge). This book is devoted to the study of ridges, the history of which is discussed in Chapter 1. Chapter 2 contains mathematical preliminaries from linear algebra, differential calculus, and tensors (algebra and calculus), which are used throughout the book. Chapter 3 contains definitions and many examples of ridges in Euclidean space. Definition. Let \(f\in C^2( \mathbb{R}^n, \mathbb{R})\). Let \(\lambda_i\), \(1\leq i\leq n\), be the eigenvalues of \(D^2f\) with \(\lambda_1\leq\dots\leq\lambda_n\). Let \(v_i\), \(1\leq i\leq n\), be corresponding unit-length eigenvectors. A point \(x\) is a \(d\)-dimensional ridge point if \(x\) is a generalized local maximum point of type \(d\) with respect to \(V=[v_1,\dots, v_{n-d}]\) (i.e., \(V^tDf(x)=0\) and \(V^tD^2f(x)V<0\) hold). Since \(V^tD^2fV= \text{diag} \{\lambda_1,\dots,\lambda_{n-d}\}\) and since the eigenvalues are ordered, the test for a ridge point reduces to \(V^tDf(x)=0\) and \(\lambda_{n-d}<0\). A point \(x\in \mathbb{R}^n\) is a \(d\)-dimensional ridge point for \(f\in C^2( \mathbb{R}^n, \mathbb{R})\) if \(\lambda_{n-d}<0\) and any of the following zero conditions holds: (1) \(P_j=0\) for \(1\leq j\leq n-d\), (2) \(\prod_{i=1}^d(D^2f-\lambda_{n-d+i}I)Df=0\), (3) \(Df\cdot\prod_{i=1}^d(D^2f-\lambda_{n-d+i}I)Df=0\). The general method for locally constructing \(d\)-dimensional ridges of \(f\in C^3\) in \(\mathbb{R}^n\) by computing the \(d\)-dimensional tangent spaces is presented. A number of detailed examples with small values of \(d\) and \(n\) are given. In Chapter 4 the natural generalization of a \(d\)-dimensional ridge point for a function on a Riemannian manifold (i.e., \( \mathbb{R}^n\) with an arbitrary positive definite metric tensor \(g_{ij}\)) is given and, for this case, all constructions from Chapter 3 are extended. In Chapter 5 the definitions of Chapter 4 are applied to finding height \(d\)-dimensional ridges of functions defined on manifolds \(M^n\) embedded in \( \mathbb{R}^p\). Section 5.2 provides an alternative definition for ridges based on principal curvatures and principal directions (the maximal curvature for ridges is motivated by the analysis of curvature properties for planar curves and for \(2-\)dimensional surfaces in \( R^3\)). Section 5.3 discusses a ridge definition which is an application of the definition of Section 5.2 to level sets. In Chapter 6, applications of ridges to image and data analysis are considered. Section 6.1 starts with the linear scale space \((x,\sigma)\in \mathbb{R}^n\times[\sigma_0,\sigma_1]\) of multiscale data \(L(x,\sigma)\) (that the vision system produces) satisfying the diffusion equation \(\sigma({\partial L}/{\partial\sigma})=\sigma^2\nabla^2L\). Using the linear scale space paradigm, a vision system measures contrast by computing gradients over a range of aperture sizes, hence a boundariness at a specified position \(x\), scale \(\sigma\), and orientation \(u\) is defined by the function \(B(x,\sigma,u)=u\cdot\sigma\nabla L(x,\sigma)\). Based on evidence that the human visual system extracts objects from images in a medial way by linking opposing boundaries, the medialness function, providing a measure of the likelihood that the point \(x\) is central to involutes at a distance \(r\), is defined as follows: \(M(x,\sigma;\rho)=-\int_{u \in S^{n-1}}B(x+ru,\sigma,u) du\), where \(\sigma=\rho t\) and \(\rho\leq 1\). The core (i.e., the points central to the object boundaries and the associated distances to the boundaries) is constructed so that it is invariant to spatial translations, spatial rotations, and zoom. Given a core with points \((x,\sigma)\), the boundary at the scale of the core (BASOC) is the boundary of the union of the solid hyperspheres of radius \(\sigma\) centered at \(x\). Since the measured spatial and scale differences are meaningful only in the context of the scale at which they are measured, the geometry of scale space is hyperbolic, it is determined by the Poincaré metric \(ds^2=(dx^2+d\sigma^2)/\sigma^2\). Then the variety of core definitions (such as height ridges of medialness, ridges of optimal scale medialness, given by separating space and scale) that have been used in medical applications is discussed. In Section 6.2, the ridges of the electron density map are considered. In Section 6.3, the ridges are applied to the analysis of vector valued-functions, the idea is to consider associated scalar functions such as divergence, squared magnitude of the curl. The author gives a general framework for computing ridges of the pressure induced by the viscous, incompressible fluid flow over a surface. In Chapter 7 the B-splines and eigensolvers are used for evaluating the function \(f\) which is given in tabular form, its derivatives, and the eigenvalues of \(D^2f\). These quantities are used for building the functions (from the rectangular lattice of data) whose zeros are the ridge points.