Computer Vision for Visual Arts Collections:

Abstract

The implementation of artificial intelligence (AI) and machine learning is becoming commonplace for visual arts libraries, archives, and museums (LAMs). In particular, computer vision, a distinct form of machine learning, has been used in arts-based LAMs to automate digital image analysis through trained algorithms to increase metadata description and collection accessibility. Linkage of LAMs with AI may seem logical, as the emerging technologies are well suited to process and analyze sizable amounts of information, such as the data held by large collection repositories. However, as interest in computer vision develops among those in the library and information science (LIS) field, there are important concerns to address. Machine-learning algorithms used in computer vision are known to reflect bias, lack transparency, and significantly impact labor. How can LAMs, as institutions dedicated to equity and access, confront these potentially harmful aspects of computer vision? Through analysis of recent case studies, accounts, and literature, this article proposes that visual arts LAMs can mitigate algorithmic bias by promoting transparency of computer vision models, demonstrating caution, and establishing accountability. The development of capable workforces in LIS through the implementation of education and collaboration is also critical to alleviate outsourcing and temporary labor. [This article is a revision of the paper that won the 2021 Gerd Muehsam Award. The award recognizes excellence in a paper written by a graduate student on a topic relevant to art librarianship or visual resources curatorship.]