[3D1-OS-22a] OS-22 (1)

[3D1-OS-22a-05] Proposal and analysis of ambiguous figures recognition method using artificial intelligence

〇Sayo Horie¹, Mori Naoki¹ (1. Osaka Prefecture University)

Keywords:Ambiguous images, Convolutional Neural Network, Kansei Engineering, Gradient-weighted Class Activation Mapping, Trompe-l'œil

Conventional research on image recognition has mainly focused on images that express unique content. For this reason, few studies have applied image recognition using artificial intelligence to ambiguous figures expressing multiple meanings, such as ambiguous figures In this research, we focus on understanding ambiguous figures by artificial intelligence.
We define three problems; the recognition problem of the ambiguous figure or not, the object identification problem of two objects in ambiguous figures, and the fixation points of ambiguous figures where artificial intelligence focused. Computer experiments were carried out taking real ambiguous figures as examples.
As a result, we succeeded in making artificial intelligence distinguish between ambiguous figures and non-ambiguous figures in ambiguous figures of human faces and landscapes. Besides, the similarity of the gazing point by artificial intelligence in portraits and ambiguous figures was confirmed by analysis using Grad-CAM.