Cartoon filter via adaptive abstraction

• We propose a non-parametric methodology to realize abstraction images.
• The redundant wavelet “à trous” algorithm is applied for details detection.
• An multi-scale circular median filter is used as a smoothing filter.
• The proposed algorithm is simple and fast on low-cost entry-level hardware.

Abstraction in computer graphics defines a procedure that discriminates the essential information that is worth keeping. Usually details, that correspond to higher frequency components, allow to distinguish otherwise similar images. Vice versa, low frequencies are related to the main information, which are larger structures. Contours themselves may also be identified by high frequencies and separate each pictured component. The underlying idea of the proposed algorithm consists in identifying these edges, by a redundant wavelet transform, and in blurring the inner areas of the components, by an adaptive circular median filter. In spite of its implementation simplicity, our unsupervised methodology provides results similar to those obtained by more complex techniques already described in the literature.