Deep learning · Computer vision
Hindi Handwriting Recognizer
Draw a Devanagari letter or digit below. A convolutional neural network, running entirely in your browser, will guess which of 46 characters it is.
- 46 classes
- 78,200 training images
- – test accuracy
- 0 servers: runs on your device
Draw here
Draw one character, for example क
Tip: draw one character, big and in the middle, with steady strokes.
Prediction
Loading model…Waiting for a drawing
The model's best guess appears here.
What the model sees. Your drawing is cropped, centred and shrunk to 32 × 32 pixels, the same format as the training images.
Characters it knows
Tap one to see what to draw. It doesn't know vowels (अ, आ…) or matras yet.
How it works
-
1
Trained on real handwriting
The model learned from 78,200 handwritten images in the UCI Devanagari Handwritten Character Dataset, with random rotations, shifts and zooms added so it copes with different writing styles.
-
2
A small CNN
Five convolution layers find strokes, curves and the top line, then two dense layers pick one of 46 classes. About 0.6 million parameters, built with PyTorch.
-
3
Runs in your browser
The trained model is exported to ONNX and runs with ONNX Runtime Web. Your drawing never leaves your device.