A handwritten OCR pipeline built for the Pulse Quest Kaggle Competition.
The project combines classical computer vision with a Convolutional Neural Network (CNN) to recognize handwritten names. Built from ground up without extenral models.
Input Image
↓
Character Segmentation (OpenCV)
↓
Character Classification (CNN)
↓
Word Reconstruction
- Connected Components Analysis
- Handcrafted filtering to remove noise
- Character extraction
- Automatic space detection
- 4-layer Convolutional Neural Network
- Trained using PyTorch
- Cross-Entropy Loss + Adam Optimizer
- Mixed Precision (AMP) training
- Character segmentation correctly matches the ground truth for 96.7% of the test dataset.
- Python
- PyTorch
- OpenCV
- NumPy
- Pandas