Generating and Evaluating Paintings that Elicit Specific Emotions with Generative Adversarial Networks

This study aimed to label paintings based on biometric data and generate paintings that elicit specific emotions using machine learning. To construct the dataset, experiments were conducted on eight participants using multi physiological sensors. Following Russell's model of emotion, the arousal-sleepy axis was assessed via skin conductance, while the pleasure-unpleasure axis was evaluated through subjective ratings. Based on these measures, paintings were categorized into four emotional states: arousal-pleasure, arousal-unpleasure, sleepy-pleasure, and sleepy-unpleasure. A generative adversarial network model was then employed to generate paintings corresponding to these emotional states. Viewing experiments were conducted using multi physiological sensors, wearable optical topography, and subjective evaluations to assess the emotional impact of the generated paintings. Skin conductance results suggested that the generated paintings may not effectively induce arousal-related emotions. Additionally, questionnaire responses revealed that paintings in the arousal-unpleasure category significantly conveyed unpleasant emotions, suggesting that machine learning can effectively capture features associated with unpleasant emotions. These findings indicate that machine learning can partially induce specific emotions through generated paintings.