Salience-adaptive painterly rendering using genetic search

We present a new non-photorealistic rendering (NPR) algorithm for rendering photographs in an impasto painterly style. We observe that most existing image-based NPR algorithms operate in a spatially local manner, typically as non-linear image filters seeking to preserve edges and other high-frequency content. By contrast, we argue that figurative artworks are salience maps, and develop a novel painting algorithm that uses a genetic algorithm (GA) to search the space of possible paintings for a given image, so approaching an "optimal" artwork in which salient detail is conserved and non-salient detail is attenuated. Differential rendering styles are also possible by varying stroke style according to the classification of salient artifacts encountered, for example edges or ridges. We demonstrate the results of our technique on a wide range of images, illustrating both the improved control over level of detail due to our salience adaptive painting approach, and the benefits gained by subsequent relaxation of the painting using the GA.