Better image search thanks to artificial intelligence

Image databases are getting bigger and bigger. In order for users to find all the images that match a search term, artificial intelligence will be used to index photos in databases. This was the goal of a project of the HTW Chur University of Applied Sciences with the Keystone photo agency.


Until now, it takes a lot of manual work to tag images in a database with appropriate keywords and additional information. Photo agencies such as Keystone keyword their own images manually so that the most suitable photos possible are specified for each search query. However, every day around 20,000 new photos are added that originate from other photo agencies but are distributed by Keystone in Switzerland.

"The problem is that the other agencies often keyword their images so that they appear as often as possible in search queries," explains Albert Weichselbraun from HTW Chur in an interview with the news agency SDA. However, users often receive a lot of images that do not necessarily match the search term. For Keystone, it has so far meant a lot of work to extract meaningful keywords from the image descriptions of the other agencies.

In future, this is to be done by an algorithm. As part of the "Imagine" project led by Weichselbraun, HTW Chur has developed machine learning methods in collaboration with Keystone to improve image searches in the Keystone database, the university announced on Tuesday.

Additional information from the web A new feature is the possibility of automatically extracting keywords from the obligatory image description text and supplementing them with additional information, Weichselbraun explained. The HTW researchers have given the system the ability to access the so-called semantic web, which offers information specially prepared for machines and enables automatic searches. Large parts of Wikipedia and data sets from the federal administration, for example, are available in this format, the university wrote.

This means that the technology can not only extract the name of a person depicted from the image description, but also add further facts from the semantic web, for example that the person is a Swiss skier. "So not only does a search for the exact name lead to the image in question, but also a broader search for Swiss skiers," Weichselbraun explained.

The project, funded by the Commission for Technology and Innovation (CTI), was completed in April. According to the HTW, the development will flow into Keystone's revised photo portal, which is to be presented to customers by the end of the year. (SDA)

More articles on the topic