Explainable AI in Art

– Visualizing Artist's Characteristics in Paintings for Enhanced Interpretation –

抄録

In recent years, deep learning has been increasingly used for automatic analysis of artistic styles and authorship in paintings. However, in practical applications, it is crucial that humans interpret and trust the model decisions. In this study, artist classification models were constructed using explainable artificial intelligence (AI) techniques, such as ResNet-50, VGG19, and Vision Transformer models, to visualize classification factors using gradient-weighted class activation mapping (Grad-CAM) and the attention mechanism. As an example, Vincent van Gogh’s masterpiece, The Starry Night was analyzed. The visualization results revealed that the regions of interest became progressively wider with the Vision Transformer, VGG19, and ResNet-50 models, in that order. This approach is expected to have practical applications, such as verification of authorship, painting style, and restoration, thereby enhancing Kansei modeling in the subjective appreciation of paintings.