Cartoon approximation with \(\alpha\)-curvelets (Q501949)

Optimal \(N\)-term approximation schemes using curvelets are presented for cartoon images, which are anisotropic and curvilinear structures. They are modeled by bivariate piecewise \(C^\beta\) functions, separated by a \(C^\beta\) discontinuity curve (with \(\beta\in(1,2]\)). A benchmark result stating that for such functions in general one cannot achieve an \(N\)-term approximation rate larger than \(\beta/2\) is given. Then the notion of \(\alpha\)-curvelet frames with \(\alpha\in [0,1]\) is introduced, where \(\alpha\) defines the degree of anisotropy and a construction of such \(\alpha\)-curvelets is presented. For \(\alpha = 0\) ridgelets are obtained, while \(\alpha=1/2\) yields second-generation curvelets. The case \(\alpha=1\) corresponds to wavelets. This curvelet construction is shown to form a tight frame for any value of \(\alpha\). For all intermediate cases, \(\alpha\in[1/2,1]\), an optimal \(N\)-term approximation (up to logarithmic factors) can be obtained for images with \(\beta=1/\alpha\).