Khmer OCR three architectures
Three recovered line OCR checkpoints: CNN + BiLSTM + CTC (step 58,000), CNN + BiGRU + CTC (step 8,500), and CNN + Transformer Encoder + CTC (step 87,000). All weights load strictly and CPU inference was executed. Checkpoints contain weights and metadata; optimizer/RNG states were removed. They are for inference and weight initialization, not exact optimizer-state resumption.
Open Compare_Models.ipynb in Google Colab and Run all. No GPU or Hugging Face account is required. It displays images, expected answers and actual predictions. Optional own-image upload is disabled by default.
Evaluation limits
The 20 demo images were selected from 40 screened printed validation examples because the Transformer predicted them correctly. Transformer 20/20, BiLSTM 14/20 and BiGRU 6/20 on this selected set do NOT establish a fair ranking or general accuracy. Stored validation CER values come from different validation procedures. Shared NFC/rounded-width preprocessing can differ from original recurrent training. Use an independent representative holdout with each model's validated preprocessing for a benchmark.
Demo source: Chanrith123333/khmer_english_ocr_image_line, validation split, CC BY 4.0. Grayscale PNG conversion and normalized labels; source row attribution is in demo/manifest.json inside demo_dataset.zip. No license is asserted here for model training data or model weights beyond rights the owner holds.
The older khmerOCR repository A1/A2/A3 approaches differ: A1 and A2 are both BiLSTM; older A3 is an attention encoder-decoder. They are not this collection's three architecture names.