GlyphReader

A deeplearning approach to classifying the ancient Egyptian hieroglyphs. The source code is written in python3 using the popular Keras framework (with a Tensorflow backend). It attempts to classify images to their Gardiner labels, such as:

Image
Gardener Label	S29	V13	G43

In addition to the source code, we also provide a dataset containing 4210 manually annotated images of Egyptian hieroglyphs found in the Pyramid of Unas. The dataset will be automatically downloaded when using train.py to train a new classifier, but is also available here

Requirements

• pip3 install numpy sklearn scipy pyyaml h5py
• tensorflow (tested with version 1.3.1)
• keras (tested with version 2.1.2)

Usage

python3 src/classify.py examples

Expected output:

Predicting the Hieroglyph type...
image name                ::: top 5 best matching hieroglyphs
200000_S29.png            --> ['S29' 'U33' 'R8' 'F12' 'Y3']
200001_V13.png            --> ['V13' 'N37' 'N18' 'V4' 'N35']
200002_V13.png            --> ['V13' 'V31' 'F22' 'N18' 'D156']
200003_G43.png            --> ['G43' 'G17' 'G21' 'W25' 'G25']
200004_D21.png            --> ['D21' 'V30' 'O50' 'D10' 'N5']
200005_O50.png            --> ['O50' 'N5' 'X6' 'D21' 'V25']
200006_X1.png             --> ['X1' 'N29' 'G1' 'D19' 'G4']
200007_M23.png            --> ['M23' 'G39' 'G25' 'I10' 'Aa26']
200008_G43.png            --> ['G43' 'G39' 'G29' 'G1' 'G4']
200009_S29.png            --> ['S29' 'Y3' 'D34' 'N5' 'W18']
200010_V13.png            --> ['V13' 'D52' 'N18' 'G17' 'F22']
200011_M23.png            --> ['M23' 'F16' 'U1' 'N14' 'M4']
200012_G43.png            --> ['G43' 'G21' 'G39' 'G1' 'G17']
200013_D21.png            --> ['D21' 'T30' 'N5' 'X6' 'U1']
200014_O50.png            --> ['O50' 'X1' 'V31' 'U33' 'U1']
200015_V13.png            --> ['V13' 'F22' 'D36' 'D46' 'V4']
200016_G43.png            --> ['G43' 'G17' 'G5' 'G7' 'G4']
200017_S29.png            --> ['S29' 'M195' 'M17' 'W18' 'M1']

Training

In case you would like to train your own classifier, use train.py. It takes no arguments, but when running it for the first time it will download the dataset, and starts training. Training itself consist of 2 phases:

1. Feature Extraction extract deeplearning features from the images (corresponding to the avg_pool layer from the InceptionV3 network).
2. Train Classifier train an SVM on the deeplearning features If you do not have a GPU, or simply want to retrain the classifier, it is possible to skip the first step and download the precomputed features directly at http://iamai.nl/downloads/features.npy, store them in intermediates/features.npy.