Visual Search

How visual search works

The system analyses every image when it is added to the library and stores a numerical description of it. A search computes the same description for the example image and returns the closest matches, ranked by a similarity score. Three kinds of description are common, often combined:

Technique	What it compares	What it finds
Perceptual hashing	A compact fingerprint of the image's structure of light and dark areas	The same image resized, recompressed, converted or lightly edited
Vision embeddings	A vector from an image recognition model that encodes what the image shows	Crops, colour variants, other photos of the same subject or scene
Colour analysis	The dominant colours and the palette	Images in a similar colour, for example a campaign palette

Perceptual hashing is precise but narrow: it finds copies of one picture. Embeddings are broader: they find pictures that resemble each other. A system that uses both can rank near-identical copies first and looser matches after them.

Visual search, reverse image search and semantic search

	Query	What is searched
Visual search in a DAM	An image	Your own library
Reverse image search on the web	An image	Public web pages, to find where an image is published
Semantic search	Words	The text attached to each asset, matched by meaning
Keyword search	Words	The exact words in the metadata

"Reverse image search" is often used for both of the first two. On the public web it is mostly used to trace where a photo appears; inside a DAM it means finding matching files in the library.

What teams use visual search for

• Finding the rest of a shoot. A designer has one product shot and wants the other angles from the same session.
• Checking before commissioning. Before booking a shoot or buying stock, check whether the library already holds a suitable image.
• Tracing an image from outside. A partner sends a screenshot from their catalogue, and visual search finds the original file with its usage rights.
• Finding an alternative. The chosen hero image is landscape and the sales deck needs a portrait image in the same style.
• Cleaning up variants. Crops, colour grades and exports of the same picture show up together, which supports duplicate detection.

Visual search in Freedam

• The camera icon in the search bar of the asset library opens visual search: drop in a reference image and get matching assets ranked by a similarity score. The reference image is not added to the library.
• Matching combines perceptual hashes of the whole image, its centre and its four quarters with a vision embedding, so crops and colour-adjusted versions are found as well as copies.
• Results only include the assets the person searching is allowed to see.
• Each image's asset page lists visually similar assets (images that share most of its AI tags) and near-duplicates (images with matching perceptual features).
• A colour picked from an image's palette finds other images with a similar colour.

Details: Find visually similar images in the Freedam docs.

Frequently asked questions

Is visual search the same as reverse image search?

The technique is the same. "Reverse image search" usually refers to web search engines that find where an image appears online. Visual search in a DAM applies the same idea to a private library.

Does visual search need tags or metadata?

No. It compares the images themselves, so it works on photos that have no title, tags or description. That makes it useful for libraries migrated from shared drives, where most files were never tagged.

Can visual search find cropped or edited versions of an image?

Usually, yes: resized, recompressed, cropped and colour-adjusted versions keep enough of the original to match. Heavy edits such as text overlays, composites or strong filters lower the score. Cropping the reference image to its main subject often improves the results.