omaressamrme/hieroglyph-classifier

Hieroglyph classifier (Gardiner sign codes)

A CNN that names a photographed Egyptian hieroglyph by its Gardiner code. Trained on 3,674 manually labelled glyph crops (50×75 px) from the Pyramid Texts of Unas (GlyphDataset, Franken & van Gemert, ACM Multimedia 2013, MIT), covering the 59 signs with at least ten examples.

Metric (held-out test glyphs, 735)	Value
Top-1 accuracy	0.969
Top-5 accuracy	0.997
Macro F1	0.939
Confidence gate ≥ 0.95	answers 90% of glyphs at 99.5% accuracy

Code, figures and the failed first run: GitHub · demo · Kaggle notebook · write-up: Teaching a CNN to read hieroglyphs · omaressam.web.app

Use

import json, numpy as np, tensorflow as tf
from PIL import Image
from huggingface_hub import hf_hub_download

model = tf.keras.models.load_model(hf_hub_download("omaressamrme/hieroglyph-classifier", "model.keras"))
labels = json.load(open(hf_hub_download("omaressamrme/hieroglyph-classifier", "labels.json")))["labels"]
img = Image.open("glyph.png").convert("L").resize((50, 75))          # width 50, height 75
x = (np.asarray(img, dtype="float32") / 255.0)[None, ..., None]      # (1, 75, 50, 1), values in [0, 1]
p = model.predict(x)[0]
print([(labels[i], round(float(p[i]), 3)) for i in p.argsort()[::-1][:5]])

Input: a single, already-cropped sign, dark strokes on a light background. The model does not detect signs in a wall photograph and does not translate.

Architecture and training

Three blocks of two 3×3 convolutions (32, 64, 128 filters) with BatchNorm, max-pooling and dropout, a Flatten layer, a 256-unit dense layer with BatchNorm and dropout 0.5, and a 59-way softmax (2.1 M parameters). Adam 1e-3 with ReduceLROnPlateau, batch 32, square-root-damped class weights, light translation/zoom augmentation with white fill, early stopping on validation accuracy (patience 20). 54 epochs, best at 34, 14 minutes on CPU. train_v2.py in this repo is the exact script; metrics.json and results/ hold every number.

Limitations

One monument, one carving style, one lighting; rare signs (fewer than 20 training glyphs) are the weak ones; 113 of the dataset's 172 codes have too few examples to be covered.

Dataset citation: M. Franken and J. C. van Gemert, "Automatic Egyptian Hieroglyph Recognition by Retrieving Images as Texts", ACM Multimedia 2013.