How to speed up image tagging

Illustration of human and robot
© artenot - #67884335

Manual tagging of images is accurate but expensive and automatic tagging is fast but still unreliable. Learn how our approach combines the best of both to implement a fast and accurate alternative.

Implementing an assisted tagging workflow

Instead of using a pure manual or automatic image tagging workflow, we have implemented a hybrid: a keyword assistant called Keyword Suggester. The special feature is, that the Keyword Suggester can improve and narrow suggestions with each user interaction. It is fully responsive and can be integrated tightly into current manual tagging approaches. Therefore the proposed workflow is lean and fast.

Automatic tag suggestions for new images

When a user uploads new images into your collection, the Keyword Suggester will automatically suggest tags. The suggestions are calculated based on the keywords of already indexed images that are visually similar to the input image. This means that the new image is used to search for similar images in the collection. Those images are already tagged. The tags will be processed and suggestions are calculated based on statistical analysis.

Screenshot of keyword suggester

The user only retrieves ready to use tag suggestions. No user interaction is required for this step and the suggestions are already complete and correct. The user just finishes the tagging process and adds the new image to the collection.

Less manual tagging effort

Manual tagging is time-consuming and means that the user has to type every single keyword. Actually, tags, that are specific to the image (for example 'New York'), should definitely be added by the user. Up to now, the automatic suggestions are more general keywords. But those are important to ensure a better retrieval rate when the image is searched by different users. It is impossible to know, which tags they are using to search for specific images. Therefore an image should be tagged with both: specific keywords and general descriptions of the image.

Fully automatic tagging systems will therefore always suffer of incomplete tagging.

Screenshot of keyword suggester

Responsive tag suggestions

As soon as a user has typed in a keyword, in this case 'New York', the suggestions will be automatically refined. In this example the Keyword Suggester automatically suggests the keyword 'Buildings' after the user added a tag manually. Every user interaction refines the tag suggestions. This way a responsive and fast workflow can be enabled: The user adds keywords as usual and the Keyword Suggester will respond in real time and produces more relevant suggestions.

Screenshot of keyword suggester

Of course, the contributor can be in charge of the tagging process and explicitly confirms suggestions by clicking on them. But it is up to you and your audience which approach you prefer: suggestions that are automatically confirmed and a contributor has to discard irrelevant tags, or the suggestions have to be confirmend explicitly before they are part of the tags.

Screenshot of keyword suggester

The Keyword Suggester is highly configurable to adapt its behaviour to your collection and specific workflow. It is even possible to not only suggest keywords but also similar images and refine the suggestions based on confirmed images. In general, we recommend to keep this refinement process as simple as possible. The Keyword Suggester is part of our product pixolution flow.