What is image hashing?

An image hash is a small fingerprint for an image.

A normal file hash, like MD5 or SHA256, answers this question:

Are these files exactly the same?

A perceptual image hash answers a more useful visual question:

Do these images look similar?

Why not just use MD5?

Imagine two screenshots that look almost the same.

Maybe one has a different timestamp. Maybe the browser rendered one pixel differently. Maybe the image was resized or lightly compressed.

With MD5, one tiny change usually creates a totally different hash.

Original image      -> 1a2b3c4d...
One pixel changed   -> 9f8e7d6c...

That is useful for exact duplicates, but useless for visual similarity.

Perceptual hashes are different

Perceptual hashes are designed so similar-looking images produce similar-looking hashes.

Original image      -> f123456789abcdef
Similar image       -> f123456789abcdee
Different image     -> 0011aa99bb44cc22

The hashes are not meant to be secret. They are meant to be compared.

Hamming distance

The difference between two hashes is called the Hamming distance.

That is just a count of how many bits are different.

Distance 0      exactly the same hash
Distance 1-5    probably very similar
Distance 6-10   possibly related
Distance 10+    often different

The exact numbers depend on the image type, hash method and your tolerance for false matches.

The idiot version

Small distance = images probably look alike.
Large distance = images probably look different.

That is the core idea.