Artificial Intelligence for content and context metadata retrieval in photographs and image groups

Digitization projects of analog photographic collections are still growing in number, and therefore such assets of images become bigger continuously. Also, there is a strong trend towards open data and interfaces to access and reuse the image resources (FAIR data). To be able to search and find images in a repository, metadata of a certain depth must be existing. Typically, indexing and valorization, done by experts that know the (photographic) collections, is necessary to achieve such meta-information. There are various metadata standards based on different concepts for the description of collections. Some, like ISAD(G), are more related to the physical structure of archives, others, like CIDOC-CRM, take into account the content of the images in detail. Enhancing the depth of indexing increases the time necessary drastically. It is also a task that is not easily scalable because specific content related knowledge is necessary. With the assistance of artificial intelligence, historic photographic collections could potentially be enhanced with metadata semi-automatically. For the successful application of machine learning, it is essential to have robust training sets. In the presented paper, we show our observations in monitoring participants indexing historic collections of photographs. In the observations of workshops of people working with photographic heritage, it was monitored how single photographs but also image groups are described. Based on that knowledge, machine learning components can be trained and optimized for that particular type of source material. The demonstrated approach has the potential to support the work of valorization substantially. In addition, the approach has, to some extent, the potential to preserve the fundamental structures of knowledge of contemporary witnesses.