armansinghh/handwriting-crnn
Building end-to-end handwritten text recognition pipeline using CRNN architecture with PyTorch, OpenCV preprocessing, and optional Streamlit deployment
What's novel
Building end-to-end handwritten text recognition pipeline using CRNN architecture with PyTorch, OpenCV preprocessing, and optional Streamlit deployment
Code Analysis
11 files read · 3 roundsA CRNN (CNN + BiLSTM + CTC) for recognizing handwritten words from the IAM dataset, with dataset preparation, training, greedy decoding, and single-image inference.
Strengths
The CRNN architecture is correctly implemented with proper spatial reduction (height→1, width preserved) and the CTC training pipeline (collate, loss, decoding) follows the standard pattern faithfully. The code is readable and the project structure is logical for its scope.
Weaknesses
The dataset loader has a dangerous infinite-loop retry pattern, the training script has an indentation bug that breaks best-model saving, and the 'tests' are just print-statement scripts with no assertions. The requirements.txt is a raw pip freeze dump rather than curated dependencies.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
11
Contributors
2
Files
10
Active weeks
3
Repository
Language
Python
Stars
2
Forks
1
License
MIT