Darayut/dbnet-khmer

DBNet Khmer Text Detector

A DBNet (Differentiable Binarization) text detection model for Khmer/English documents — MobileNetV3 backbone + FPN + DBHead. Detects bounding boxes around text lines; it does not perform OCR/recognition.

Trained from scratch on synthetic Cambodian-style documents (plain docs, ID cards, official letters). See dbnet-text-detector for the full training pipeline, data generators, and inference code.

2026-07-31 — the default weights changed. khmer_dbnet_mobilenetv3.onnx now serves v2, a fine-tune on realistic ID-card/official-letter documents. It detects noticeably more text on real photographed documents than v1 did. The previous weights remain available unchanged as khmer_dbnet_mobilenetv3_v1.onnx — pin that file if you need the old behaviour.

Files

File	Description
khmer_dbnet_mobilenetv3.onnx	Default. Currently identical to v2. Recommended for inference (torch-free, via onnxruntime)
khmer_dbnet_mobilenetv3.pth	PyTorch checkpoint of the default — for fine-tuning or DBNet.pytorch-based inference
khmer_dbnet_mobilenetv3_v2.onnx / .pth	v2 explicitly, pinned
khmer_dbnet_mobilenetv3_v1.onnx / .pth	v1 explicitly, pinned — the original release

Usage

pip install dbnet-text-detector

from dbnet_detector import DBNetDetector

detector = DBNetDetector.from_pretrained()          # current default (v2)
boxes, scores = detector.detect("your_image.jpg")

detector = DBNetDetector.from_pretrained("v1")      # pin an older version
detector = DBNetDetector.from_pretrained(preset="display")   # large display text

Weights are downloaded from this repo on first use and cached locally (~/.cache/huggingface/hub).

Thresholds matter more than you'd expect

The library defaults to the document preset (box_thresh=0.3, short_size=1088), which is tuned for dense scanned or photographed pages. The older box_thresh=0.6, short_size=736 values silently dropped whole text lines on real documents — on one sample page, 24 detected lines versus 79 with the current defaults.

Large display text (posters, captions over photos) is the opposite case and needs downscaling — use preset="display" (short_size=320). The model was trained on small text lines inside 640px crops, so very large glyphs fall outside its training scale.

Versions & metrics

Version	Training	Val F1
v1	from scratch on plain synthetic docs	0.987 (v1 val set)
v2 (default)	v1 fine-tuned on synthetic ID cards / official letters	0.933 (v2 val set)

⚠️ These two numbers are not comparable and do not rank the models. Each is measured on its own in-distribution synthetic validation set, so each says "this model learned its own data" — not that v1 is better. On real photographed documents v2 detects substantially more text, which is what the fine-tune was for. Both figures are synthetic and in-distribution; neither is evidence of generalization to arbitrary real-world photographs. Validate on your own images.

License

MIT