Prof. Dr. Jürgen Döllner

Flow-based Image and Video Abstraction

This project investigates automatic non-photorealistic image processing techniques for the creation of simplified stylistic illustrations from color images, videos and 3D renderings based on the bilateral and difference of Gaussians filters.

Our method extends the approach of [Winnemöller et al. 2006] to use iterated bilateral filtering for abstraction and difference-of-Gaussians (DoG) for edge extraction by adapting it to the local orientation of the input. To represent local orientation we construct a smooth tensor field. From the eigenvectors of this tensor field we derive a vector field that has similar characteristics as the edge tangent flow (ETF) of [Kang et al. 2007], but its computation is much less expensive. Besides gradient calculation, only smoothing with a box or Gaussian filter is necessary. In contrast to that, ETF construction requires several iterations of a nonlinear filter with large filter kernel.

The xy-separated bilateral filter [Pham and van Vliet 2005] used by Winnemöller et al. suffers from horizontal and vertical artifacts. These artifacts appear in particular when the filter is applied iteratively. Our approach works by first filtering in direction of the gradient and then filtering the intermediate result in perpendicular direction. When applied iteratively our approach does not suffer from horizontal or vertical artifacts and creates smooth output at curved boundaries.

DoG edges often look frayed and don’t reassemble straight line and curve segments very well. To work around this limitation, [Kang et al. 2007] recently introduced the concept of flow-based difference-of-Gaussians which, compared to DoG edges, create more coherent lines. They replaced the DoG filter by a flow-guided anisotropic kernel whose shape is defined by the ETF. Comparable high-quality results can be achieved by a separated implementation with corresponding reduced computational complexity. We first apply a one-dimensional difference-of- Gaussian filter in direction of the gradient and then apply smoothing along the vector field that we derive from the smoothed structure tensor.


XDoG: An eXtended difference-of-Gaussians compendium including advanced image stylization

Winnemöller, Holger; Kyprianidis, Jan Eric; Olsen, Sven C. in Computers & Graphics 2012 .

Recent extensions to the standard Difference-of-Gaussians (DoG) edge detection operator have rendered it less susceptible to noise and increased its aesthetic appeal. Despite these advances, the technical subtleties and stylistic potential of the DoG operator are often overlooked. This paper offers a detailed review of the DoG operator and its extensions, highlighting useful relationships to other image processing techniques. It also presents many new results spanning a variety of styles, including pencil-shading, pastel, hatching, and woodcut. Additionally, we demonstrate a range of subtle artistic effects, such as ghosting, speed-lines, negative edges, indication, and abstraction, all of which are obtained using an extended DoG formulation, or slight modifications thereof. In all cases, the visual quality achieved by the extended DoG operator is comparable to or better than those of systems dedicated to a single style.
Further Information
Tags cgs