Unfourtunately, I couldn’t find a source stating it would be required. AFAIK it’s been assumed that they would use perceptual hashes, since that’s what various companies have been suggesting/presenting. Like Apple’s NeuralHash, which was reverse engineered. It’s also the only somewhat practical solution, since exact matches would be easily be circumvented by changing one pixel or mirroring the image.

Patrick Breyer’s page on Chat Control has a lot of general information about the EU’s proposal.

Matched using perceptual hash algorithms that have an accuracy between 20% and 40%.

base12 has the advantage of being divisible by 2, 3, 4 and 6, while base10 is only divisible by 2 and 5.