handwritten-meitei-mayek-recognition

Handwritten Meitei Mayek character recognition

Six networks fine-tuned on the 55-class TUMMHCD dataset of isolated handwritten Meitei Mayek characters: ConvNeXt-T, EfficientNetV2-S and ResNet-50-D, all ImageNet-pretrained, each trained once on the image alone and once with five size features (the _meta files). Their average gets 98.12% top-1 accuracy on the 12,794 test images.

Use

git clone https://github.com/chingkheinganba231005/Handwritten-Meitei-Mayek-Recognition
cd Handwritten-Meitei-Mayek-Recognition && pip install -r requirements.txt
huggingface-cli download Chingkheinganba/handwritten-meitei-mayek-recognition --local-dir models

from mayek.recognizer import Recognizer

rec = Recognizer("models")
preds, seen = rec.predict("photo.jpg", source="upload")   # or source="canvas" for drawings
print(preds[0].cls.display, preds[0].cls.name, preds[0].prob)

Inputs are brought to the scale of the dataset's 24 x 24 scans before the usual preprocessing, so drawings and photos work, but accuracy on handwriting unlike the TUMMHCD scans will be lower than the test figure.

Demo: huggingface.co/spaces/Chingkheinganba/handwritten-meitei-mayek-recognition. Training code and every experiment: https://github.com/chingkheinganba231005/Handwritten-Meitei-Mayek-Recognition.

Dataset: D. Hijam and S. Saharia, "On developing complete character set Meitei Mayek handwritten character database", The Visual Computer 38, 525–539 (2022).