IsaakAlemu/amharic-character-recognition

# Amharic Handwritten Character Recognition A neural network that classifies handwritten Amharic characters, built to practice neural network fundamentals (forward propagation, dense layers, softmax classification) while tackling a real, underserved problem: Amharic handwriting recognition, an area with far less tooling and research attention than Latin-script OCR. ## Why this scope Amharic's full script (Ge'ez/Fidel) has 33 base consonants, each written in 7 vowel-modified forms ("orders") — 231+ total character shapes. This project deliberately scopes down to the **33 base (first-order) characters only**. This keeps the model, a plain dense neural network, appropriately matched to the problem's complexity, while still tackling a real classification task grounded in actual research datasets rather than a toy problem like MNIST. ## Results | Model | Test Accuracy | |---|---| | Logistic Regression (baseline) | 82.4% | | Dense Neural Network | 86.3% | | **Convolutional Neural Network (CNN)** | **95.8%** | The dense neural network achieved an 86.3% test accuracy (~3.9 percentage points over the linear baseline). Upgrading to a CNN yielded a substantial leap to **95.77% test accuracy** (+9.5 pp over Dense NN, +13.4 pp over baseline). Across 3 runs with different random seeds, CNN test accuracy consistently ranged from **95.4% to 96.5%** (averaging **95.89%**), confirming the stability of spatial feature learning across initializations. ### Training curves Validation accuracy tracks *above* training accuracy for most of training — a signature of the data augmentation (random rotation/translation/zoom) working as intended: it makes the training task harder, which reduces overfitting rather than causing it. ### Confusion matrix Predictions are strongly concentrated on the diagonal (correct), with errors sharply reduced across all character classes. **The primary remaining confusion:** ደ (`169de`, `U+12F0`) predicted as ጸ (`211Te`, `U+1338`): - **Dense NN** …