A review of bandlet methods for geometrical image representation (Q2641449)

It is well known that for a \(C^\alpha\) image \(f\) defined on the unit square, say, its best \(M\)-term wavelet approximation \(f_M\) satisfies \(\| f-f_M\| ^2=O(M^{-\alpha})\) in \(L^2\)-norm when \(C^\alpha\) orthogonal wavelets are used. Such estimates fail, though, for discontinuous images. Bandlets are a type of wavelet adapted to contours and motivated by the following observation: though an image might be discontinuous across a contour, it can still be smooth in a direction parallel to the tangent of the contour. The bandlet transform exploits such anisotropic regularity by constructing orthogonal vectors that are elongated in the direction where the function has maximal regularity. This observation is motivated, in turn, by triangularizations in finite element decompositions in which anisotropic triangles are used to focus on regions of high variation in numerical solutions of PDEs. Ordinary wavelet thresholding in this context corresponds to approximation using elements with square supports of size inversely proportional to magnitude of signal variation. Orthogonal bandlet bases are obtained from a wavelet basis by means of a cascade of orthogonal operators parameterized by local geometry. Here, one observes that wavelet coefficients of \(f\) are integer samples \(f_j(2^jn)\) in which \(f_j(x)=f\ast \psi_j\) where \(\psi_j(x)=2^{-j}\psi(-2^{-j} x)\). If a contour in the unit square locally satisfies \(x_2=\gamma(x_1)\) then one defines bandlet coefficients in terms of wavelet coefficients along a flow determined by a vector field extension \((1,\widetilde{\gamma}'(x_1))\) of the tangents to \((x_1,\gamma(x_1))\). Here, one defines \((\widetilde{x}_1, \widetilde{x}_2)= w(x_1,x_2)= (x_1,x_2- \widetilde{\gamma}(x_1))\) by setting \(\widetilde{f}_j(\widetilde{x})= f_j(w^{-1}(\widetilde{x}))\). The precise mechanics of the orthogonal bandletization make use of a piecewise Alpert wavelet approximation. Different choices of flows give rise to different bandlet bases among which one can choose a best basis according to the criterion of minimizing mean squared error based on, for example, using a small number of terms above a given threshold. The details of defining the orthogonal bandlet bases and choosing a best basis are provided. Applications to denoising and image restoration are discussed in detail. These ideas are further refined with the notion of grouplets that group wavelet coefficients along multiscale association fields analogous to quadtrees.