Novelty and Self-Modification in the Context of Figurative Expression-Based Evolutionary Art

Abstract

Evolutionary Art combines evolutionary computation approaches and computer graphics in order to generate artworks. Within this field, it has been shown that it is possible to evolve interesting artworks that can please the human eye. However, most of these artworks are created with human supervision and the large majority of artworks generated are abstract. In contrast, this dissertation is inspired on a previous work that aims to evolve figurative images, that is, images that resemble some object to the human eye, without any human supervision during the process. Instead, the supervision process resorts to an object classifier, trained to recognize specific objects, which is used to assign fitness. In this dissertation, this work is expanded in order to demonstrate that, depending on the object classifiers created, it is possible to generate any object. Furthermore, this dissertation explores the evolution of ambiguous images, that is, images that can resemble several objects at the same time within the same region, using a set of classifiers to evaluate instead of just one. Finally, this dissertation also focuses on the evolution of images that can be considered as di↵erent as possible among them, using novelty search techniques. Experimental results show that it is possible to generate images that are (1) figurative and (2) ambiguous, from a computational and human perspective. The results also indicate that the classifiers’ robustness plays an important role in approximating the computational and the human point of view. Furthermore, the success in employing novelty mechanisms depends also on the classifiers used: when a single classifier is used to guide evolution, novelty search tends to result in a broader set of diverse images. However, when several classifiers are used, novelty search is only able to promote diversity when the classifiers are permissive.