Hieroglyph classifier
A CNN that names an Egyptian sign by its Gardiner code
Trained on 3,674 hand-labelled glyphs cut from the Pyramid Texts of Unas, it reads 96.9% of held-out glyphs correctly at the first guess and 99.7% within five, across the 59 signs with at least ten examples. With a confidence gate at 0.95 it answers 90% of glyphs on its own at 99.5% accuracy and leaves the rest to a person. The first training run overfitted to under 50%; the published run fixed the head, the class weights and the augmentation, and both are in the repo.
GitHub codeHugging Face modelHugging Face live demoKaggle notebookdev.to article
Data: GlyphDataset, Franken & van Gemert, ACM Multimedia 2013 (MIT).
-
D21 mouth -
G43 quail chick -
M17 reed leaf -
N35 ripple of water -
O50 threshing floor -
S29 folded cloth -
X1 bread loaf
one dot per sign · size = test glyphs · the dashed line is 20 training images · every sign below F1 0.8 sits left of it
