Automatically Classifying Art Images Using Computer Vision

Abstract

Millions of art images have been digitized over the last several decades. This has created new opportunities for art scholars and historians. However, searching and navigating these art images is difficult because of the sparsity of the metadata and contextual information used to describe these images. Unless one knows the exact title and artist, finding related paintings is a difficult task without the metadata. The research in this project addresses this challenge by developing unsupervised computer vision methods that will extract metadata automatically from paintings. Our dataset will include more than 2 million art images from Artstor, a non-profit organization that distributes art images to libraries and universities. If successful, we plan to build an interactive interface for exploring the extracted features and for developing a recommender system that could be used on platforms such as Artstor.