Portfolio Project
Handwriting Legibility Scoring
PyTorch CNN Fine-Tuning
Context
My wife says my handwriting is hard to read. I wanted an objective score.
Approach
- Built three digit-recognition models (simple baseline → CNN).
- Trained on MNIST (60,000 digits) and selected the best model.
- Deployed it behind a serverless scoring API for the live demo.
Impact
- The best model reached 99.1% accuracy on MNIST.
- My digits scored 72.5% legible; 0, 3, 5, and 8 were the toughest.
- My wife was right.
Experiment Setup
I used MNIST as a baseline, then scored my own handwriting with the same setup.
- Dataset: MNIST (60,000 train / 10,000 test) for digit recognition.
- Goal: compare model types and quantify my own legibility with a consistent metric.
- Output: per-digit accuracy and a confusion matrix to see which digits I write most ambiguously.
Model Iteration
- Model 1: simple linear classifier as a baseline.
- Model 2: small CNN (TinyVGG-style) to capture strokes and curves.
- Model 3: deeper CNN (VGG16-style) with the best test accuracy (99.1%).
Custom Handwriting Evaluation
- Built a loader so new images can be preprocessed and scored consistently.
- Computed overall legibility (72.5%) and found my hardest digits (0, 3, 5, and 8).
- Used the breakdown to make it actionable (which digits to practice).
What I'd Improve
- Train on EMNIST or other real handwriting datasets to reduce domain shift from MNIST.
- Add augmentation (stroke thickness, rotation, blur) to better match camera/scanner variation.
- Calibrate confidence so the model can say 'uncertain' instead of forcing a guess.
