V. Sundaram, A. Coyle
Southern Methodist University,
Keywords: machine learning, shallow neural networks, art, style
Summary:For centuries, art history has grappled with the problem of authorship. In the absence of secure documentation, art historians have relied on the practice of connoisseurship, or the attribution of artworks to artists based on the close study of visual and material similarities between works. For our project, we applied machine learning to the age-old problem of attribution using a dataset of paintings from the Renaissance through the 20th century taken from ImageNet, an image database used in advancing computer vision and deep learning research. Our approach centered on using the pre-trained ImageNet layers as feature extractors, allowing us to extract features from paintings such as background strokes and styles. We then used this data to train a neural network with a set of dense layers to identify the artists with their artwork. Our neural network correctly identified the artist with his/her artwork with an accuracy of 85-90%. Whereas many neural networks focus on the identification of objects within pictures, our algorithm sought to understand the style of the picture, thus identifying the general characteristics of the artwork. By focusing on style, our model was trained using three of the dense layers of the neural network rather than the entire deep layer structure often used in object detection, using the pre-trained patterns from ImageNet architecture. The results are promising for further investigations into the application of artificial intelligence and machine learning in the area of visual arts. However, beyond the problem of art attribution, our approach could be applicable in fields such as medicine, ecology, and computer security where the concept of “style” could be generalized to a higher-order problem space rather than a focus on discrete, identifiable problem elements.