What is a perceptual hash? How apps spot similar images

How does a perceptual hash work?

The common average-hash and difference-hash methods follow the same pattern:

1. Shrink the image to a tiny grid, often 8x8 or 9x8 pixels, which removes size and fine detail.
2. Convert it to greyscale so colour shifts do not matter.
3. Compare each pixel with the average (aHash) or with its neighbour (dHash).
4. Write a 1 or a 0 for each comparison, giving a 64-bit fingerprint.

pHash adds a discrete cosine transform before the comparison, which makes it steadier under compression and small brightness changes. Two images are compared by counting how many of the 64 bits differ, the Hamming distance. A distance of a handful of bits usually means the same picture.

Because the fingerprint is tiny, comparing thousands of photos is fast: a phone can check every pair in a large library in seconds without opening the full images again.

How is it different from a file hash?

A cryptographic file hash is designed so that changing one byte changes the whole hash, the avalanche effect. That is perfect for proving two files are identical and useless for asking whether they look alike. Perceptual hashing deliberately does the opposite: it is locality-sensitive, so close inputs give close outputs.

Where do perceptual hashes fall short?

Basic hashes struggle with crops, rotations, mirrored images and overlays, and they can mark unrelated pictures with similar layouts as matches, such as two plain sky shots. Research also shows hashes can be fooled on purpose with tiny pixel changes. Large-scale systems such as Microsoft's PhotoDNA use more robust designs for those reasons.

Modern photo cleaners often go further, using image embeddings from a neural network to compare content, plus capture time to keep separate moments apart. That is how a tool can group burst frames while leaving last year's shot of the same view alone. See near-duplicate photo for what these groups look like in practice.

Colour is also invisible to the basic methods. Two photos with the same layout but different colours, such as a red and a blue car parked in the same spot, can hash almost alike, which is why tools confirm matches with a second check.