A convolutional neural network that classifies handwritten letters (A–Z) from uploaded photos. Built with PyTorch, trained on EMNIST, and served through a Streamlit web app. Includes a custom OpenCV preprocessing pipeline that turns real-world photos (shadows, uneven lighting, arbitrary framing) into the clean format the model expects.
Upload a photo of a single handwritten letter and the app preprocesses it, runs it through a trained CNN, and returns the predicted letter with confidence scores. If the model isn't confident the image is a clear letter, it says so rather than guessing.
Model: A CNN (two convolutional layers + two fully-connected layers) trained on the EMNIST "letters" split (~124k training images, 26 classes). Reaches ~93% test accuracy.
Preprocessing pipeline (the part that handles real photos):
- Convert to grayscale
- Adaptive thresholding — handles uneven lighting and shadows by comparing each pixel to its local neighborhood rather than a global cutoff
- Contour detection — locates the letter and rejects shadow-edge artifacts
- Crop, square-pad, and center to match EMNIST's framing
- Resize to 28×28 and apply EMNIST's orientation and normalization
- Test vs. real-world gap: ~93% on clean EMNIST test data; lower on real uploaded photos, since real handwriting/lighting differs from training data.
- Ambiguous letters: The model confuses genuinely similar pairs —
i/l,g/q,u/v— because these are nearly identical as isolated characters. This is a data-ceiling limit, not a tuning issue. - No "not a letter" detection: Shown a non-letter, the model still picks a letter, though usually with low confidence — the app flags low-confidence predictions for this reason.
- Input sensitivity: Works best on a single, clearly-written letter that fills the frame, with even lighting and minimal background.
pip install -r requirements.txt
streamlit run app.py
app.py— Streamlit interface + preprocessing pipelinetraining.py— model definition and trainingemnist_cnn.pth— trained model weights
