Building a Handwriting Recognition System
summary
Built a sketch box canvas that rasterizes strokes, applies a 2‑second lock to wait for stroke completion, and sends binarized, anti‑aliased bitmaps to on‑device models. Tried classical OpenCV + KNN filtering, then the MLKit digital ink recognizer, then a custom TensorFlow Lite CNN trained in Colab on ~60,000 MNIST/EMNIST samples expanded via augmentation to ~2 million, yielding ~98% train accuracy and a ~100KB on‑device model that successfully recognizes letters.
AI-written summary · about
chapters (16)
- 0:00Handwriting recognition project
- 1:01Project motivation
- 3:53Android input mechanism
- 7:45ML integration overview
- 10:14Edge detection approach
- 14:31Testing MLKit handwriting
- 16:17TensorFlow Lite models
- 17:21Training with MNIST
- 21:55Convolutional neural network architecture
- 24:01Data augmentation
- 27:31Extending to letters
- 30:40How gradient descent works
- 32:04Cross-validation tuning
- 34:10Model performance demo
- 35:32Next steps
- 38:18Q&A on languages
related talks
25mBuilding a Handwriting Recognition System for the New York Times Crossword
44mBuilding Handwriting Recognition System for the NY Times Crossword
44mHello Deep Learning
49mExploring Neural Networks for Recognition of Hand-drawn Shapes
21mRecognizing Human features using Deep Networks
44mAI for every developer
19mReacTex: using React Native and Neural Networks to recognize handwritten equations
57mWriting a Neural Net from Scratch