rajesh-1902/greek-htr-crnn-ctc

Greek Sentence-Level HTR - CRNN + CTC (Improved)

Performance

Metric	Before	After
CER	7.81%	7.60%
WER	36.42%	35.81%

Improvements Applied

• Post-processing to fix space errors around punctuation
• Fine-tuned at LR=1e-5 for additional epochs

Architecture

• CNN: 5-block with BatchNorm (64->128->256->256->512)
• RNN: 2x BiLSTM hidden=512 with Dropout(0.3)
• Decoder: CTC (BestPath / BeamSearch)
• Input: 1024x64, Max sequence: 128 chars

Usage

python main.py --infer --image your_image.png
python main.py --infer --image your_image.png --beamsearch