Annotations as Knowledge Practices in Image Archives: Application of Linked Open Usable Data and Machine Learning

Abstract

We reflect on some of the preliminary findings of the Participatory Knowledge Practices in Analogue and Digital Image Archives (PIA) research project around annotations of photographic archives from the Swiss Society for Folklore Studies (SSFS) as knowledge practices, the underlying technological decisions, and their impact. The aim is not only to seek more information but to find new approaches of understanding the way in which people’s memory relate to the collective, public form of archival memory and ultimately how users figure in and shape the digital archive. We provide a proof-of-concept workflow based on automatically generated annotations comprising 53,481 photos that were subjected to object detection using Faster R-CNN Inception ResNet V2. Of the detected objects, 184,609 have a detection score greater than 0.5, 123,529 have a score greater than 0.75, and 88,442 have a score greater than 0.9. A threshold of 0.75 was set for the dissemination of our annotations, compatible with the W3C Web Annotation Data Model (WADM) and embedded in our IIIF Manifests. In the near future, the workflow will be upgraded to allow for the co-existence of various, and occasionally conflicting, assertions made by both human and machine users. We believe that Linked Open Usable Data (LOUD) standards should be used to improve the sustainability of such an ecosystem and to foster collaboration between actors in cultural heritage.