Perceptual hashing calculates a short signature from the visual features of media rather than from the raw file data. Traditional hashes such as MD5 or SHA-256 change completely if a single byte changes; perceptual hashes are designed so that two visually similar images produce the same or very similar values.
Common methods include average hash (comparing the brightness of small blocks to the overall average), difference hash (comparing neighbouring pixels), and wavelet hash (analysing frequency components). They tolerate common changes such as recompression, format conversion, and resizing, though heavy cropping or editing can break a match.
Privly uses perceptual hashing as part of content fingerprinting, so copies of your content can be recognised even when they've been recompressed or resized, and takedowns can be filed against the matches.