Geometry-driven diffusion in computer vision (Q1900217)

This book is the fruit of a European-American transatlantic collaboration, sponsored by an ESPRIT-NSF grant (travel and subsistence) awarded for the period 1993-1996. The book brings together the important groups active in this field, many of which have an outstanding record of achievements. A short synopsis of the book: Chps. 1 and 2 by Lindeberg and ter Haar Romeny give a tutorial introduction to linear scale-space. From Chp. 3 onwards the book focuses on non-linear or geometry-driven diffusion. Coarsely, two major approaches can be discriminated, each with their variations: the variational approach, were the energy of a functional is minimised, the functional being some cost function that can be manipulated, and on the other hand the nonlinear PDE approach, where the evolution of the image is expressed as some function of invariant properties of the image. Historically geometry-driven diffusion was introduced as being some local function of edge-strength. This is discussed in Chp. 3 by Perona, Shiota and Malik, together with a critical analysis of the resulting non-linear PDE and the discrete maximum principle. In Chp. 4 by Whitaker and Gerig not only edges are involved, but a complete `feature space' gives rise to multivalued or vector-valued diffusion, using both (higher order) features in a geometry-limited diffusion scheme, as local frequency decomposition in a spectra-limited diffusion scheme. The variational approach is given a firm Bayesian basis by Mumford, one of its original authors, in Chp. 5. Four probabilistic models are presented from reasoning and axioms, and shown on examples. The variational approach is a prototypical example of a `free- discontinuity problem'. Critical questions like uniqueness of the solution and existence of an optimal segmentation are raised by Leaci and Solimini in Chp. 6. Edges get displaced when they are blurred. Nordström in Chp. 7 studies edges as finite unions of smooth curves (line drawings) and demonstrates the existence of a solution in this particular case. In the next two chapters a link is being established between the variational approach and the PDE-based evolutions. In Chp. 8 Richardson and Mitter extend the variational approach with the theory of \(\Gamma\)- convergence, i.e. the approximation of one functional by another. Replacing the edge set with a smooth function then leads to edge detection by coupled partial differential equations. These are further discussed and elaborated in Chp. 9 by Proesmans, Pauwels and van Gool, extending the applications to second order smoothing, multispectral images, optic flow and stereo. An attempt to some unification is presented in Chp. 10 by Alvarez and Morel, who summarize the morphological approach to geometry-driven diffusion in the light of many other `scale-space' or multi-scale approaches. The authors establish a number of basic principles for the visual pyramid and derive the `fundamental equation' of shape analysis. The general form of the nonlinear diffusion equations are always some function of a particular invariant. In Chp. 11 Olver, Sapiro and Tannenbaum first give a tutorial on basic invariant theory, starting from Lie groups and prolongations. Images can be considered embedded sets of curves. The authors then describe in detail the invariant curve evolution flows under Euclidean, similarity, affine and projective transformations, as well as the impact of certain conservation laws. Kimia, Tannenbaum and Zucker take methods from optimal control theory in Chp. 12. Dynamic programming and the application of the Hamiltonian- Jacobi equation are applied to shape theory, morphology, optic flow, non- linear scale-space and shape-from-shading. Two chapters introduce and apply the notion of the covariant formalism for coordinate free reasoning. Florack et al. study in Chp. 13 a special nonlinear scale-space, which can be mapped onto a linear scale-space. The Perona and Malik equation is studied in this context, and a log-polar foveal scale-space is presented as one of the examples. Eberly in Chp. 14 gives detailed examples of nonlinear scale-space properties once a proper metric is chosen. The appendix of Chp. 13 is a tutorial for the covariant formalism. The volume ends with a chapter on implementations.