Ambiguous figures recognition and generation by Vision Transformer and Generative Adversarial Networks

Abstract

There are few researches that have applied artificial intelligence to generate images such as trompe l'oeil which can be interpreted in multiple ways. In this study, for the purpose of understanding and generating ambiguous figures by artificial intelligence, we conducted numerical experiments on the identification and generation of ambiguous figures by acquiring ambiguous figure-like characteristics from images generated by the image identification model and generation model. As a result, we succeeded in showing that the generated image is a ambiguous figure by visualizing the basis for the decision that the artificial intelligence identified it as a ambiguous figure. The effectiveness of the proposed method is demonstrated by confirming that the method for generating ambiguous figures can be extended to the methods for generating landscapes and portraits.