khmer-ocr-selected-demo

Expanded model specific gallery

Screened 500 unique printed validation images using the three saved models. Selected exact matches: Transformer 380, BiLSTM 273, BiGRU 70. 49 images were correct for all three. These sets overlap; there are 426 selected unique images in total. These are selected demonstration examples, not overall accuracy or a fair benchmark.

Open expanded live demo in Colab. Run all, choose a model and image, then click Run real prediction. CPU is sufficient.

Download the gallery and image folders. Unzip it and open Gallery.html. Choose a model, or choose Correct for all three models. The all_three folder contains the 49 common exact-match images; individual model folders include more examples and answers.csv.

Images are cropped printed text lines. New images, handwriting and full pages can behave differently. Labels were kept from the source dataset and normalized consistently, not replaced with predictions. expanded_screening_results.json includes failures. The earlier 20-image demo below remains available.

Selected successful Khmer OCR demo examples

20 cropped text-line images verified with Architecture 3 best.pt (step 87,000). These examples were selected after screening 40 images from the source validation split. This dataset is for demonstrating inference, not measuring general accuracy. Full predictions, including failures, are in screening_results.json. No claim of writer-independent testing or absence of overlap across source corpora is made.

Source: Chanrith123333/khmer_english_ocr_image_line, validation rows 0–39, CC BY 4.0. Images converted to grayscale PNG; labels normalized to NFC with whitespace collapsed. Attribution remains in each manifest row.

Unzip demo_dataset.zip and run the GitHub demo. Model: https://huggingface.co/Seoul12/khmer-ocr-architecture3-best