Exploring the Space of Abstract Textures by Principles and Random Sampling

Exemplar-based texture synthesis methods try to emulate textures observed in our visual world. Yet the field of all possible textures (natural or not) has been little explored. Indeed, existing abstract synthesis methods focus on a single generation rule and generate a rather limited set of textures. This limitation can be overcome by combining randomly various generation principles and rule parameters. Doing so gives access to a vast and still unexplored set of possible images. In this paper, we introduce an image sampling method combining the main painting techniques of abstract art. This sampler synthesizes what we call multi-layered textures. The underlying image model extends three abstract image synthesis models: the dead leaves model, the spot noise, and fractal generators. By respecting minimal self-similarity rules keeping Gestalt theory grouping principles at each texture layer, the abstract textures remain understandable to human perception. The complexity of the generated textures derives from the systematic and randomized use of shape interaction principles taken from abstract art such as occlusion, transparency, exclusion, inclusion, and tessellation.