Automatic Tagging

It is also called auto-tagging, AI tagging or automated tagging. In a digital asset management system it usually runs as one step of ingestion, next to preview generation and text recognition.

Why it matters

Manual tagging does not keep up with volume. Every image needs someone to look at it and type a set of terms, and a library that receives thousands of images a month needs a lot of someone's time just to describe them. In practice that work is done partly or not at all, and untagged images can only be found by browsing folders.

Automatic tagging gives every image a baseline description the moment it arrives. It does not replace the fields only your organisation can fill in, but it means no image enters the library with nothing to search on.

What an AI tagging model produces

Current models, which combine image recognition and a language model, typically return:

• Tags: a list of short terms such as kitchen towel, herbs, linen, flat lay, green.
• A caption: one or two sentences describing the image.
• Text in the image: the words printed on packaging, posters or signs, read by OCR.
• A classification: which kind of asset this is (product shot, lifestyle photo, logo, document), chosen from a list the organisation defines.

Older systems returned only a list of labels with a confidence score for each. Models that return captions and read text give search much more to work with, because people often search with a phrase or a product name rather than a single object word.

Where automatic tagging goes wrong

• It does not know your vocabulary. A model can see a sofa. It cannot know that the sofa is called "Nordwik 3-seater" in your catalogue, belongs to the autumn campaign, or may only be used in Europe. Product names, campaigns, rights and approval status stay human or rule-based work.
• It is generic. Tags such as "indoor" or "object" match too many images to help. Good systems instruct the model to avoid them.
• It can be wrong. Similar-looking products, unusual angles and abstract images produce wrong tags. Tags need to be easy to correct.
• Its language may differ from yours. Many models tag in English by default. In a multilingual organisation, check how the system handles search in other languages.
• It changes over time. A newer model tags the same image differently. Re-tagging a whole library can overwrite corrections made by people.

How a DAM should handle it

When you evaluate a DAM, check these points with your own images:

1. Tags are generated at upload without any manual step, for every image, whichever way it enters the library.
2. AI tags are kept separate from the tags and fields people enter, so you can tell them apart and filter on either.
3. People can correct or remove AI tags on one asset and in bulk.
4. AI tags, captions and text read from images are all searched by keyword search and by semantic search.
5. Your own fields, keywords and controlled vocabularies are never overwritten by AI output.
6. You can see which model is used and what the tagging costs.

Automatic tagging in Freedam

In Freedam, every image, EPS file and single-page PDF is analysed when it is uploaded, imported from a connected source or sent through the API. One analysis returns 15 to 20 tags, a caption, the text found in the image and the asset class. The results appear on the asset's AI tab, apart from the keywords and metadata fields that people fill in, which the analysis never changes. They are searchable by keyword, by meaning, and through the AI Tags, AI Caption and OCR Text filters. Administrators choose the AI model and see its cost on the AI Settings page.

Details: Automatic Tagging in the Freedam docs.

Frequently asked questions

How accurate is AI image tagging?

Current models describe common objects, settings and colours reliably on clear photos. They are less reliable on products that look alike, unusual angles and abstract images, and they cannot know internal names such as product lines or campaigns. Treat AI tags as a starting point that people can correct.

What is the difference between auto-tagging and manual tagging?

Manual tagging is done by a person, usually with the organisation's own terms, and covers what only the organisation knows: product names, campaigns, rights. Auto-tagging is done by a model on every image and covers what can be seen. Most libraries use both.

Does auto-tagging work on videos and documents?

It depends on the system. Many tag images only. For a document, the text it contains is usually more useful for search than tags. For a video, a transcript of the speech is.