Cartoon Extraction Based on Anisotropic Image Classification

We propose a new approach for the extraction of cartoons from 2D aerial images. Particularly in city areas, these images are mainly characterized by rectangular geometries of locally varying orientation. The presented method is based on a joint classification of the shape orientation and a rectangular structure preserving prior in the restoration of image shapes. Mathematically, an anisotropic area functional encodes the preference for edges aligned to locally preferable directions and a higher order regularization term ensures a smooth variation of these directions. The concrete model is an anisotropic version of the Rudin-OsherFatemi (ROF) scheme with a position dependent anisotropy. Given the knowledge of the anisotropic image structure, the restoration process can be significantly improved, in particular the round-off effect of the ROF model can be reduced. By combining the extraction of the anisotropy with the denoising method in a joint variational approach, we obtain a suitable classification method, in which a tedious direct anisotropy estimation can be avoided. The implementation is based on a finite element discretization and an energy minimization via a stepsize-controlled gradient method. Instructive synthetic images are considered to demonstrate the methods performance and the approach is applied to aerial images as a prototype application.