Draw or upload a handwritten number. A neural network running in your browser reads it, digit by digit.
Loading model…
Write a number here, leaving small gaps between digits
Predicted number
Nothing yet
What the model saw
White = ink after thresholding. Boxes = the pieces sent to the network.
How it works
Segmentation: Otsu thresholding separates ink from paper, connected-component analysis finds each blob, and overlapping or wide blobs are merged or split.
Classification: each piece is resized to 28×28 and read by a convolutional neural network, first trained on MNIST and then fine-tuned on real handwritten numbers.
Join: the digits are combined from left to right. Predictions under 60% confidence are flagged as uncertain.
89.35% of numbers recognised exactly on 2,000 held-out test images. Known limitation: digits that touch are read as one piece.