Genetic Programming for Non-Photorealistic Rendering

by Maryam Baniasadi (mary_baniasadi@hotmail.com)
Supervisor: Brian Ross

This research uses genetic programming to produce non-photorealistic (NPR) images. A number of major technical enhancements are introduced here. A multi-objective sum of ranks scoring strategy is used, which produces results that satisfy the majority of fitness objectives, while minimizing outliers solutions common with the more conventional Pareto ranking. Another innovation is that colour mixing expression is fully evolved by GP. This permits a wide range of different classical and innovative NPR effects to arise. The GP language is also extended with a library of colour rendering building blocks, for example, functions such as darken, burn, dodge, and others. A few basic paint rendering algorithms are introduced, which apply paint brush strokes to the canvas in an economical fashion. This results in different rudimentary "painting styles", which will highlight the painterly effects of the evolved NPR expressions. An assortment of aesthetic image analysis procedures are used in a multi-objective fashion, resulting in a fully automated evolutionary art system.

The galleries below show hand-selected images from many different GP runs. A variety of different styles of NPR effects will be seen, ranging from natural media styles (watercolour, oil, acrylic, "drippy" paint) to innovative effects (mixed media, "chemical reaction" paintings,...).

Publications

Examples

Images copyright (c) 2013 Maryam Baniasadi.


Back up: http://www.cosc.brocku.ca/~bross/