Promised too much? AI in use in the image archive of the ETH Library


METADATA ONLY
Loading...

Author / Producer

Date

2023-11-28

Publication Type

Conference Paper

ETH Bibliography

yes

Citations

Altmetric
METADATA ONLY

Data

Rights / License

Abstract

With 3.6 million photographs – from the period between 1860 and today, the Image Archive of the ETH Library in Zurich is one of the largest historical image archives in Switzerland. Since 2021, the Image Archive uses artificial intelligence for content-based image indexing, in addition to intellectual indexing, and for the translation of metadata. Users benefit from additional research options. The automatically generated tags are closer to everyday language and go into more detail. The Image Archive uses autotagging in a complementary way: It does not replace the intellectual work with the subject headings but complements it. Although the autotags produced good results on average, a quality check was initiated. After analysing 75% of the autotags, the deletion rate was around 50 to 70%. While the computing time is considerable, the financial effort is negligible. In view of the initial processing, which took over 16 months, the question arose as to whether the computational effort required to generate the autotags was worthwhile at all. An online survey provided insights into the acceptance of the autotags on the one hand and the research techniques of the users on the other.

Publication status

published

External links

Book title

EVA Berlin 2023: Elektronische Medien & Kunst, Kultur und Historie. Konferenzband | Proceedings

Journal / series

Volume

Pages / Article No.

295 - 301

Publisher

BTU Brandenburgische Technische Universität

Event

EVA Berlin 2023: Electronic Media & Visual Arts

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Organisational unit

00060 - Abt. ETH-Bibliothek / ETH-Bibliothek check_circle

Notes

Conference lecture held on December 1, 2023.

Funding

Related publications and datasets

Is documented by: