OpenThai2.0-OCR-Text

Fast Thai + English printed-text OCR that runs on a CPU. It is the engine behind iApp Document OCR v3.0, released as open weights with its runtime.

• Detector: PP-OCRv6-tiny (1.8 MB).
• Recognizer: a Thai CTC line recognizer trained by iApp (2.4 MB).
• Runtime: ONNX models on OpenVINO (or ONNX Runtime), CPU only. About 0.5 CPU-seconds per dense A4 page, roughly 950 pages a minute on 8 Xeon cores.

It reads a page image and returns its text in reading order, with a box and a confidence for every line. It does not write tables or Markdown; for that, see OpenThai2.0-OCR-Markdown.

Model at a glance

Parameters	1.02M in all: detector 428K (PP-OCRv6-tiny), recognizer 587K
Files	det.onnx 1.8 MB and rec.onnx 2.4 MB, float32; OpenVINO runs them in bfloat16 on CPUs with AMX
Recognizer	CTC over text lines 48 pixels high; 530 output classes from a 528-character Thai and English dictionary
Output	the page's text in reading order, with a box and a confidence for every line
Languages	Thai and English, printed
Hardware	CPU only; no GPU needed
Speed	0.39 s per dense A4 page, one page at a time; 0.5 CPU-seconds a page, about 950 pages a minute on 8 Xeon cores
Memory	about 450 MB of RAM per worker process
Base	PaddlePaddle PP-OCRv6-tiny: the detector as shipped, the recognizer fine-tuned from it with a new Thai output layer
Training data	410,270 real text lines (court scans, parliament budget PDFs, Wikisource old books) and synthetic lines
Licence	Apache-2.0
Version	v1.0.0, November 2026

How it compares

Every engine read the same pages and was scored the same way. The score is character accuracy, 100 × (1 − CER), so higher is better, against checked gold, after markup is stripped from both sides; reading order counts. A page an engine fails on counts as empty.

The 12 printed sets: parliament budget documents and Royal Gazette pages (each as a hand-verified set and as a larger set), SEA-DocBench Thai, Wikisource old books, Administrative Court judgments (50 documents), ThaiOCRBench full pages, th-docs-tables text-only pages, and three OCR104 sets (document crops, Thai+English crops, document photos). The table shows six of them, with handwriting, which is not in the average.

Engine	Average, 12 printed sets	Court judgments	Royal Gazette	Budget (verified)	Old books	ThaiOCRBench	Handwriting
OpenThai2.0-OCR-Text (this model)	87.8	97.0	96.3	99.5	77.1	82.0	10.3
OpenThai2.0-OCR-Markdown (our page-to-Markdown model)	90.3	96.2	98.0	98.2	79.7	90.6	78.9
Typhoon OCR V2 2B	89.6	96.2	94.6	97.3	77.2	87.3	76.7
Typhoon OCR 1.5 2B	87.5	92.4	92.5	97.2	78.6	71.4	94.5
Gemini 2.5 Flash	87.1	94.7	92.6	81.0	77.9	88.8	86.7
dots.mocr	86.4	96.3	93.1	99.5	76.3	78.8	39.4
OneOCR (Microsoft, Windows)	85.0	95.1	96.6	99.8	78.6	81.8	91.0
Google Cloud Vision	83.6	93.9	87.8	75.0	73.6	85.6	60.8
PaddleOCR 3 PP-OCRv5 (th)	81.1	90.4	82.9	93.1	55.5	75.8	41.6
Qwen3.5-9B	80.1	34.5	88.2	91.8	68.5	82.0	54.1
NVIDIA Nemotron 3.5 Parse (default: boxes, classes, Markdown)	89.0	96.9	92.2	98.9	76.9	92.0	57.3
NVIDIA Nemotron 3.5 Parse (plain-text prompt)	77.3	82.9	78.8	83.0	71.7	84.7	41.1
Tesseract 5 (tha+eng)	75.4	71.7	82.3	71.3	52.2	75.0	26.4
PaddleOCR-VL 1.6	72.4	80.8	79.7	70.3	57.6	68.2	13.4
DeepSeek-OCR-2	69.0	92.8	84.7	91.4	55.2	77.5	3.9

Bold marks the best score in each column. The average is over the 12 printed sets listed above the table; handwriting is not in it.

Speed

Engine	Runs on	Seconds per page	CPU seconds per page
OpenThai2.0-OCR-Text (this model)	CPU	0.39	0.504
OpenThai2.0-OCR-Markdown (our page-to-Markdown model) ‡	GPU (H100)	2.97	—
Typhoon OCR V2 2B †	GPU (H100)	16.72	—
Typhoon OCR 1.5 2B	GPU (H100)	2.79	—
Gemini 2.5 Flash	API	3.33	—
dots.mocr	GPU (H100)	4.78	—
OneOCR (Microsoft, Windows)	CPU	0.72	2.22
Google Cloud Vision	API	0.99	—
PaddleOCR 3 PP-OCRv5 (th)	CPU	15.20	16.291
Qwen3.5-9B	GPU (H100)	3.81	—
NVIDIA Nemotron 3.5 Parse (default: boxes, classes, Markdown) ‡	GPU (H100)	3.04	—
NVIDIA Nemotron 3.5 Parse (plain-text prompt) ‡	GPU (H100)	3.99	—
Tesseract 5 (tha+eng)	CPU	1.40	1.662
PaddleOCR-VL 1.6 §	GPU (H100)	1.71	—
DeepSeek-OCR-2	GPU (H100)	1.81	—

Bold marks the fastest in each column.

One page at a time, the engine alone on its GPU or API. Most engines read the same 24 printed pages from the sets above. ‡ Nemotron 3.5 Parse and OpenThai2.0-OCR-Markdown: 24 court-judgment pages (full A4 pages of dense text). † Typhoon OCR V2 2B: the 65 th-docs-tables pages, served with --enforce-eager, which it needs on our vLLM version. § PaddleOCR-VL 1.6: every sample page, one at a time. CPU seconds are one single-threaded worker on an Intel Xeon 8468 core; at 0.504 CPU seconds a page, this model reads about 950 pages a minute on 8 cores. On shorter pages the GPU models are faster: OpenThai2.0-OCR-Markdown's own card gives 1.6 s a page, the median of 12 table pages.

Use

from runtime.line_engine import LineOCR      # needs openvino (or onnxruntime), opencv-python-headless, pyclipper, numpy

ocr = LineOCR(".", threads=1)                # this folder: det.onnx, rec.onnx, rec_dict.txt, config.json
r = ocr.read("page.png")                     # path, bytes, BGR ndarray or PIL image
print(r["text"])                             # lines in reading order (columns one after another)
r["lines"]                                   # [{"text", "confidence", "row", "box"}]; boxes in the input image's coordinates
r["rotation"], r["text_angle"]               # quarter turns applied (0/90/180/270) and the page's text angle in degrees

For throughput, run one worker process per core with threads=1. Threads in one process stop scaling at about 2 cores, because the Python parts hold the GIL.

What the runtime handles

• Page orientation from the text itself: sideways and upside-down pages are turned the way their lines read best.
• Deskew above 2.5°, then a second detection; boxes come back in the caller's coordinates.
• Reading order: two- and three-column pages are read column by column; tables, budget pages and lists numbered across the row stay in row order.
• Colour: each column's background is subtracted, so light-on-dark text, coloured slides and gradients read.
• Clean-up: detached-mark fragments are dropped, lines merged across a gutter are split, and Thai/Latin look-alikes are fixed (a lone W glued to Thai becomes พ).

Robustness suite: 22 generated conditions (rotation, tilt, huge and tiny text, colour, faded print, photocopy noise, phone photos, tables, columns), 20 held-out pages each, bars fixed before the first run. All 22 pass.

Files

File	What
det.onnx, det.yml	Text-line detector: PaddlePaddle PP-OCRv6-tiny, Apache-2.0, as shipped
rec.onnx, rec_dict.txt	Thai CTC line recognizer. Input [N, 3, 48, W]; output softmax [N, W/8, 530]
config.json	Line filter (drop_below 0.8, junk_below 0.9, junk_invalid 0.05, min_chars 2) and crop handling (color: flatten)
pytorch/rec_weights.pt	Recognizer weights for further training (transformers PPOCRV6TinyRecForTextRecognition, classifier replaced)
runtime/line_engine.py, runtime/thai_norm.py	The runtime

Training

The recognizer starts from the PP-OCRv6-tiny recognizer (Apache-2.0) with a new Thai output layer, fine-tuned with CTC on:

• synthetic Thai/English lines rendered on the fly from 85 OFL/TLWG fonts plus 29 bold weights, with court-text, dictionary, numeric and field strings, English from wikitext, short labels and cross-script look-alike drills;
• 261,877 Administrative Court scan lines, labelled with the court's own DOC text;
• 41,101 parliament budget-PDF lines, text-layer labels kept only where the recognizer independently agrees;
• 107,292 Wikisource old-book lines (level-3 pages; the level-4 pages are a test set).

No OneOCR output and no Gemini labels were used, no personal-data sets, and every benchmark document is excluded by document.

Acknowledgements

The recognizer was trained on NVIDIA H100 GPUs provided by Siam AI. The detector and the recognizer's starting point are PaddlePaddle's PP-OCRv6-tiny.

Limitations

• Printed text only. Handwriting and scene text are weak (see the table); use a VLM for those.
• Decorative display fonts: short labels in heavily stylised fonts can be misread.
• Stamps and seals come out as text, as they are text on the page.
• Court judgments are in-domain: court scans (other documents than the 150 tested) are in the training data.
• The bf16 speed needs AMX (Sapphire Rapids or later); OpenVINO falls back to fp32 on other CPUs.

PP-OCRv6 models © PaddlePaddle authors, Apache-2.0; see NOTICE.