Painterly Image Processing Techniques

Watercolor paintings are widely known for the diversity of applications and effects that can be produced with them (see Figure ) . Since the simulation of all these drawing effects would be laborious, Curtis et al. [#!CONF_CG_INTERACTIVE-1!#] use a physical shallow-water model to simulate the distribution of pigments.

Figure: Watercolor effects - (a) dry-brush (b) edge-darkening (c) back-run (d) granulation (e) flow pattern (f) glazing

A system that produces visible brush strokes of various size to generate impressionistic painting style was presented by Hertzmann [#!EVL-1998-433!#].

Figure: A painting produced by variable brush strokes and sizes
Figure: A shower door effect

The author uses multiple passes over the image in order to give the image a more 'refined' look. Furthermore, strokes are rendered in random order so as to prevent an undesired appearance of regularity. From a source image, an artistic approximation is generated by progressively painting strokes of decreasing size over each other. The areas in which the strokes should be placed are derived from blurred versions of the source image. The resulting painting style vary depending on the user's specification (see Figure ).

The application of the approach to video can be found in [#!SYMP_NPAR-2000-1!#,#!hertzmann01relax!#] as well as [#!litwinowicz97video!#]. In all of these papers, frame-to-frame coherence is an important subject. Because the brush strokes stick to the image plane rather than on the painted objects, a shower door appearance is the consequence (see Figure ).

There are dozens of other approaches that produce paintings in various styles. These techniques, however, are tailored specificly for that sole purpose. It would be nice if the computer could learn by itself how a specific painting style works, which would be learning-by-example. A recent paper that shows how this can be accomplished is the one on ``Image Analogies'' by Hertzmann et al. [#!EVL-2001-121!#]. It uses example images to produce all kinds of image filters. The whole process of filtering can be broken up into two stages:

1. the design phase: the program is trained with both the filtered and the original version of an image. This is repeated with different images, and the resulting filter is stored.
2. the application phase: the filter is applied to a target image.

More precisely, the problem the approach tries to solve is: For a given unfiltered training image A, a filtered training image A' and an unfiltered target image B, compute the filtered target image B' (see Figure ) .

Figure: The target image is filtered like the training images

Applications of this technique include

• traditional image filters such as blur, sharpen, etc.
• texture synthesis: the production of new textures from an image
• super-resolution: a higher-resolution image is obtained from a lower-resolution source
• texturization with an arbitrary image artistic filters and painting styles (e.g. impressionism, oil painting etc., see Fig. )
• texture-by-numbers: areas of a source image are annotated with colors. Through a simple painting interface, the artist now has the opportunity to create new images which resemble the old one (Figure ).

Figure: From a real landscape image, a new one is created by specifying the areas that should be copied into the target image