Commit ·
64e20ae
0
Parent(s):
AksonOCR-Lite 1.0.0 models
Browse files- .gitattributes +49 -0
- BENCHMARKS.md +104 -0
- LICENSE +201 -0
- MODEL_CARD.md +57 -0
- NOTICE +26 -0
- README.md +253 -0
- assets/AksonOCR-logo.svg +13 -0
- assets/charts/accuracy_light.png +3 -0
- assets/charts/accuracy_mdp_light.png +3 -0
- assets/charts/accuracy_sea_light.png +3 -0
- assets/charts/accuracy_thai_light.png +3 -0
- assets/charts/rec_light.png +3 -0
- assets/charts/speed_accuracy_light.png +3 -0
- assets/charts/speed_light.png +3 -0
- assets/charts/speed_mdp_light.png +3 -0
- assets/charts/speed_sea_light.png +3 -0
- assets/charts/speed_thai_light.png +3 -0
- assets/viz/cluster_tokens.gif +3 -0
- assets/viz/demo1.gif +3 -0
- assets/viz/sample1.gif +3 -0
- assets/viz/sample2.gif +3 -0
- bench/RESULTS.md +67 -0
- bench/SPEED.md +113 -0
- config.json +34 -0
- det.onnx +3 -0
- dewarp_grid.onnx +3 -0
- dewarp_mask.onnx +3 -0
- doc_ori.onnx +3 -0
- layout.onnx +3 -0
- rec.onnx +3 -0
- rec_charset.json +1 -0
- tatr.onnx +3 -0
- textline_ori.onnx +3 -0
.gitattributes
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
| 2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
| 3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
| 7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
| 8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
| 10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
| 11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
| 12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
| 13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
| 14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
| 15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
| 16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
| 17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
| 18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
| 19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
| 20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
| 21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
| 22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
| 23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
| 24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
| 25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
| 27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
| 28 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
| 29 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
| 30 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
| 31 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
| 32 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
| 33 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
assets/charts/accuracy_light.png filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
assets/charts/accuracy_mdp_light.png filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
assets/charts/accuracy_sea_light.png filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
assets/charts/accuracy_thai_light.png filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
assets/charts/rec_light.png filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
assets/charts/speed_accuracy_light.png filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
assets/charts/speed_light.png filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
assets/charts/speed_mdp_light.png filter=lfs diff=lfs merge=lfs -text
|
| 44 |
+
assets/charts/speed_sea_light.png filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
assets/charts/speed_thai_light.png filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
assets/viz/cluster_tokens.gif filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
assets/viz/demo1.gif filter=lfs diff=lfs merge=lfs -text
|
| 48 |
+
assets/viz/sample1.gif filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
assets/viz/sample2.gif filter=lfs diff=lfs merge=lfs -text
|
BENCHMARKS.md
ADDED
|
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# AksonOCR-Lite benchmarks
|
| 2 |
+
|
| 3 |
+
[← back to README](README.md)
|
| 4 |
+
|
| 5 |
+
All numbers on this page: AksonOCR-Lite against 12 other Thai OCR systems (Typhoon OCR V2/1.5, OpenThai2.0,
|
| 6 |
+
Nemotron 3.5 Parse, dots.mocr, Chandra OCR 2, Surya OCR 2.5, Indra OCR, Gemini 4 Hydrogen, Claude Haiku 5.5), each
|
| 7 |
+
run as its authors recommend and scored by the same official scorers. How every system was run, and every column:
|
| 8 |
+
[bench/RESULTS.md](bench/RESULTS.md) (accuracy) and [bench/SPEED.md](bench/SPEED.md) (speed). Scripts:
|
| 9 |
+
[`bench/`](https://github.com/Remtir/AksonOCR-Lite/tree/main/bench). Logos (from each maker's Hugging Face profile) are trademarks of their owners and only identify
|
| 10 |
+
the system.
|
| 11 |
+
|
| 12 |
+
- [AksonOCR-Lite: default vs `--dewarp`](#aksonocr-lite-default-vs---dewarp)
|
| 13 |
+
- [Accuracy](#accuracy): average, MDPBench-th, ThaiOCRBench DP, SEA-DocBench-th
|
| 14 |
+
- [Speed](#speed): average, per benchmark
|
| 15 |
+
- [Text recognition only](#text-recognition-only)
|
| 16 |
+
|
| 17 |
+
## AksonOCR-Lite: default vs `--dewarp`
|
| 18 |
+
|
| 19 |
+
**MDPBench, Thai (160 pages: 40 digital, 120 photographed)**, official scorer:
|
| 20 |
+
|
| 21 |
+
| Setting | Overall ↑ | Digital ↑ | Photo ↑ |
|
| 22 |
+
|---|---:|---:|---:|
|
| 23 |
+
| `--dewarp` (recommended for photos) | **85.6** | 89.5 | 84.3 |
|
| 24 |
+
| default (no dewarp) | 79.5 | 89.5 | 76.2 |
|
| 25 |
+
|
| 26 |
+
With `--dewarp`, pages that already fill the image (scans, born-digital) are left unwarped; only photos with
|
| 27 |
+
background around the page are flattened.
|
| 28 |
+
|
| 29 |
+
| Benchmark | Metric | default | `--dewarp` |
|
| 30 |
+
|---|---|---:|---:|
|
| 31 |
+
| [ThaiOCRBench](https://github.com/scb-10x/ThaiOCRBench) — Document parsing (211 samples) | STEDS ↑ | 0.640 | **0.641** |
|
| 32 |
+
| [SEA-DocBench](https://huggingface.co/datasets/xingranzhao/SEA-Vision), Thai (1,506 pages) | text NED ↓ | **0.0580** | 0.0595 |
|
| 33 |
+
| SEA-DocBench, Thai | table TEDS ↑ | 0.749 | **0.757** |
|
| 34 |
+
|
| 35 |
+
## Accuracy
|
| 36 |
+
|
| 37 |
+
### Average of the three benchmarks
|
| 38 |
+
|
| 39 |
+
Three scores on a 0–100 scale, averaged: MDPBench-th (overall: 40 digital + 120 photographed pages), ThaiOCRBench Document parsing (×100; the better
|
| 40 |
+
of its two modes, item question / page parse) and SEA-DocBench-th text accuracy (100 − NED×100).
|
| 41 |
+
|
| 42 |
+

|
| 43 |
+
|
| 44 |
+
### MDPBench-th
|
| 45 |
+
|
| 46 |
+

|
| 47 |
+
|
| 48 |
+
### ThaiOCRBench, Document parsing
|
| 49 |
+
|
| 50 |
+

|
| 51 |
+
|
| 52 |
+
### SEA-DocBench-th
|
| 53 |
+
|
| 54 |
+

|
| 55 |
+
|
| 56 |
+
Speed vs accuracy as a scatter plot: [speed_accuracy_light.png](assets/charts/speed_accuracy_light.png).
|
| 57 |
+
|
| 58 |
+
## Speed
|
| 59 |
+
|
| 60 |
+
Default settings (no dewarp), timed end to end on each benchmark's pages: MDPBench-th 160 pages, ThaiOCRBench DP 211,
|
| 61 |
+
SEA-DocBench-th 151 (every 10th page). One at a time: AksonOCR-Lite 1 process × 16 threads, VLMs 1 request.
|
| 62 |
+
Batched: AksonOCR-Lite 16 single-thread processes on the CPU, VLMs vLLM with up to 256 requests in flight on one
|
| 63 |
+
RTX 5090. CPU: EPYC 7763.
|
| 64 |
+
|
| 65 |
+
### Average of the three benchmarks
|
| 66 |
+
|
| 67 |
+

|
| 68 |
+
|
| 69 |
+
| System | Hardware | s / page, one at a time | pages / s, batched |
|
| 70 |
+
|---|---|---:|---:|
|
| 71 |
+
| **AksonOCR-Lite** | CPU only (EPYC 7763) | **1.48** | **1.76** |
|
| 72 |
+
| **AksonOCR-Lite** | RTX 5090 (`--device cuda`) + CPU recognizer | **1.08** | – |
|
| 73 |
+
| OpenThai2.0-OCR-Markdown | RTX 5090, vLLM | 4.14 | 4.88 |
|
| 74 |
+
| Typhoon OCR V2 0.8B | RTX 5090, vLLM | 4.17 | 2.14 |
|
| 75 |
+
| Typhoon OCR V2 2B | RTX 5090, vLLM | 5.11 | 1.23 |
|
| 76 |
+
|
| 77 |
+
### MDPBench-th
|
| 78 |
+
|
| 79 |
+

|
| 80 |
+
|
| 81 |
+
### ThaiOCRBench, Document parsing
|
| 82 |
+
|
| 83 |
+

|
| 84 |
+
|
| 85 |
+
### SEA-DocBench-th
|
| 86 |
+
|
| 87 |
+

|
| 88 |
+
|
| 89 |
+
Per-benchmark tables, median/p90, per-stage split and token rates: [bench/SPEED.md](bench/SPEED.md).
|
| 90 |
+
|
| 91 |
+
## Text recognition only
|
| 92 |
+
|
| 93 |
+
The recognizer alone against the other Thai CPU line recognizer,
|
| 94 |
+
[OpenThai2.0-OCR-Text](https://huggingface.co/iapp/OpenThai2.0-OCR-Text), on the same 469 pre-cut line crops
|
| 95 |
+
(internal printed-line set, 184 document + 285 photo lines), no detection or layout, CPU only
|
| 96 |
+
([`bench/rec_bench.py`](https://github.com/Remtir/AksonOCR-Lite/blob/main/bench/rec_bench.py)). AksonOCR-Lite here is the recognizer without its 180° re-read, which is
|
| 97 |
+
meant for upside-down lines in full pages; with it: [bench/SPEED.md](bench/SPEED.md#text-recognition-only).
|
| 98 |
+
|
| 99 |
+

|
| 100 |
+
|
| 101 |
+
| Recognizer | exact match % ↑ | character error (NED×100) ↓ | ms / line, 1 thread ↓ | lines / s, 16 processes ↑ |
|
| 102 |
+
|---|---:|---:|---:|---:|
|
| 103 |
+
| **AksonOCR-Lite** | **88.6** | **1.15** | **8.1** | **1,800** |
|
| 104 |
+
| OpenThai2.0-OCR-Text (OpenVINO) | 38.1 | 12.10 | 10.9 | 1,187 |
|
LICENSE
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Apache License
|
| 2 |
+
Version 2.0, January 2004
|
| 3 |
+
http://www.apache.org/licenses/
|
| 4 |
+
|
| 5 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 6 |
+
|
| 7 |
+
1. Definitions.
|
| 8 |
+
|
| 9 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 10 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 11 |
+
|
| 12 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 13 |
+
the copyright owner that is granting the License.
|
| 14 |
+
|
| 15 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 16 |
+
other entities that control, are controlled by, or are under common
|
| 17 |
+
control with that entity. For the purposes of this definition,
|
| 18 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 19 |
+
direction or management of such entity, whether by contract or
|
| 20 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 21 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 22 |
+
|
| 23 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 24 |
+
exercising permissions granted by this License.
|
| 25 |
+
|
| 26 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 27 |
+
including but not limited to software source code, documentation
|
| 28 |
+
source, and configuration files.
|
| 29 |
+
|
| 30 |
+
"Object" form shall mean any form resulting from mechanical
|
| 31 |
+
transformation or translation of a Source form, including but
|
| 32 |
+
not limited to compiled object code, generated documentation,
|
| 33 |
+
and conversions to other media types.
|
| 34 |
+
|
| 35 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 36 |
+
Object form, made available under the License, as indicated by a
|
| 37 |
+
copyright notice that is included in or attached to the work
|
| 38 |
+
(an example is provided in the Appendix below).
|
| 39 |
+
|
| 40 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 41 |
+
form, that is based on (or derived from) the Work and for which the
|
| 42 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 43 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 44 |
+
of this License, Derivative Works shall not include works that remain
|
| 45 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 46 |
+
the Work and Derivative Works thereof.
|
| 47 |
+
|
| 48 |
+
"Contribution" shall mean any work of authorship, including
|
| 49 |
+
the original version of the Work and any modifications or additions
|
| 50 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 51 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 52 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 53 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 54 |
+
means any form of electronic, verbal, or written communication sent
|
| 55 |
+
to the Licensor or its representatives, including but not limited to
|
| 56 |
+
communication on electronic mailing lists, source code control systems,
|
| 57 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 58 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 59 |
+
excluding communication that is conspicuously marked or otherwise
|
| 60 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 61 |
+
|
| 62 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 63 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 64 |
+
subsequently incorporated within the Work.
|
| 65 |
+
|
| 66 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 67 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 68 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 69 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 70 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 71 |
+
Work and such Derivative Works in Source or Object form.
|
| 72 |
+
|
| 73 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 74 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 75 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 76 |
+
(except as stated in this section) patent license to make, have made,
|
| 77 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 78 |
+
where such license applies only to those patent claims licensable
|
| 79 |
+
by such Contributor that are necessarily infringed by their
|
| 80 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 81 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 82 |
+
institute patent litigation against any entity (including a
|
| 83 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 84 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 85 |
+
or contributory patent infringement, then any patent licenses
|
| 86 |
+
granted to You under this License for that Work shall terminate
|
| 87 |
+
as of the date such litigation is filed.
|
| 88 |
+
|
| 89 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 90 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 91 |
+
modifications, and in Source or Object form, provided that You
|
| 92 |
+
meet the following conditions:
|
| 93 |
+
|
| 94 |
+
(a) You must give any other recipients of the Work or
|
| 95 |
+
Derivative Works a copy of this License; and
|
| 96 |
+
|
| 97 |
+
(b) You must cause any modified files to carry prominent notices
|
| 98 |
+
stating that You changed the files; and
|
| 99 |
+
|
| 100 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 101 |
+
that You distribute, all copyright, patent, trademark, and
|
| 102 |
+
attribution notices from the Source form of the Work,
|
| 103 |
+
excluding those notices that do not pertain to any part of
|
| 104 |
+
the Derivative Works; and
|
| 105 |
+
|
| 106 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 107 |
+
distribution, then any Derivative Works that You distribute must
|
| 108 |
+
include a readable copy of the attribution notices contained
|
| 109 |
+
within such NOTICE file, excluding those notices that do not
|
| 110 |
+
pertain to any part of the Derivative Works, in at least one
|
| 111 |
+
of the following places: within a NOTICE text file distributed
|
| 112 |
+
as part of the Derivative Works; within the Source form or
|
| 113 |
+
documentation, if provided along with the Derivative Works; or,
|
| 114 |
+
within a display generated by the Derivative Works, if and
|
| 115 |
+
wherever such third-party notices normally appear. The contents
|
| 116 |
+
of the NOTICE file are for informational purposes only and
|
| 117 |
+
do not modify the License. You may add Your own attribution
|
| 118 |
+
notices within Derivative Works that You distribute, alongside
|
| 119 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 120 |
+
that such additional attribution notices cannot be construed
|
| 121 |
+
as modifying the License.
|
| 122 |
+
|
| 123 |
+
You may add Your own copyright statement to Your modifications and
|
| 124 |
+
may provide additional or different license terms and conditions
|
| 125 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 126 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 127 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 128 |
+
the conditions stated in this License.
|
| 129 |
+
|
| 130 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 131 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 132 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 133 |
+
this License, without any additional terms or conditions.
|
| 134 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 135 |
+
the terms of any separate license agreement you may have executed
|
| 136 |
+
with Licensor regarding such Contributions.
|
| 137 |
+
|
| 138 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 139 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 140 |
+
except as required for reasonable and customary use in describing the
|
| 141 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 142 |
+
|
| 143 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 144 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 145 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 146 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 147 |
+
implied, including, without limitation, any warranties or conditions
|
| 148 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 149 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 150 |
+
appropriateness of using or redistributing the Work and assume any
|
| 151 |
+
risks associated with Your exercise of permissions under this License.
|
| 152 |
+
|
| 153 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 154 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 155 |
+
unless required by applicable law (such as deliberate and grossly
|
| 156 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 157 |
+
liable to You for damages, including any direct, indirect, special,
|
| 158 |
+
incidental, or consequential damages of any character arising as a
|
| 159 |
+
result of this License or out of the use or inability to use the
|
| 160 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 161 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 162 |
+
other commercial damages or losses), even if such Contributor
|
| 163 |
+
has been advised of the possibility of such damages.
|
| 164 |
+
|
| 165 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 166 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 167 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 168 |
+
or other liability obligations and/or rights consistent with this
|
| 169 |
+
License. However, in accepting such obligations, You may act only
|
| 170 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 171 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 172 |
+
defend, and hold each Contributor harmless for any liability
|
| 173 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 174 |
+
of your accepting any such warranty or additional liability.
|
| 175 |
+
|
| 176 |
+
END OF TERMS AND CONDITIONS
|
| 177 |
+
|
| 178 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 179 |
+
|
| 180 |
+
To apply the Apache License to your work, attach the following
|
| 181 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 182 |
+
replaced with your own identifying information. (Don't include
|
| 183 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 184 |
+
comment syntax for the file format. We also recommend that a
|
| 185 |
+
file or class name and description of purpose be included on the
|
| 186 |
+
same "printed page" as the copyright notice for easier
|
| 187 |
+
identification within third-party archives.
|
| 188 |
+
|
| 189 |
+
Copyright 2026 AksonOCR-Lite contributors
|
| 190 |
+
|
| 191 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 192 |
+
you may not use this file except in compliance with the License.
|
| 193 |
+
You may obtain a copy of the License at
|
| 194 |
+
|
| 195 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 196 |
+
|
| 197 |
+
Unless required by applicable law or agreed to in writing, software
|
| 198 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 199 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 200 |
+
See the License for the specific language governing permissions and
|
| 201 |
+
limitations under the License.
|
MODEL_CARD.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- th
|
| 5 |
+
- en
|
| 6 |
+
library_name: onnx
|
| 7 |
+
pipeline_tag: image-to-text
|
| 8 |
+
tags:
|
| 9 |
+
- ocr
|
| 10 |
+
- thai
|
| 11 |
+
- document-parsing
|
| 12 |
+
- text-recognition
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# AksonOCR-Lite models
|
| 16 |
+
|
| 17 |
+
Weights used by the [AksonOCR-Lite](https://github.com/Remtir/AksonOCR-Lite) package
|
| 18 |
+
(`pip install aksonocr-lite`). Everything here runs on ONNX Runtime; nothing needs PyTorch or Paddle.
|
| 19 |
+
|
| 20 |
+
| File | What it is | Origin | License |
|
| 21 |
+
|---|---|---|---|
|
| 22 |
+
| `rec.onnx`, `rec_charset.json` | Thai + English text-line recognizer (ours) | fine-tuned from PaddlePaddle `PP-OCRv6_tiny_rec` | Apache-2.0 |
|
| 23 |
+
| `tatr.onnx` | table structure (rows, columns, headers, spans) | `microsoft/table-transformer-structure-recognition-v1.1-all`, converted | MIT |
|
| 24 |
+
| `dewarp_mask.onnx`, `dewarp_grid.onnx` | page segmentation + dewarping grid (optional) | MonkeyOCRv2 document preprocessor, converted | Apache-2.0 |
|
| 25 |
+
| `layout.onnx`, `det.onnx`, `doc_ori.onnx`, `textline_ori.onnx` | layout, text detection, page / line orientation | PaddlePaddle `PP-DocLayoutV3`, `PP-OCRv6_medium_det`, `PP-LCNet_x1_0_doc_ori` / `textline_ori` ONNX exports, unmodified copies | Apache-2.0 |
|
| 26 |
+
|
| 27 |
+
## Recognizer (`rec.onnx`)
|
| 28 |
+
|
| 29 |
+
- **Architecture:** PP-OCRv6 tiny (PPLCNetV4 backbone + SVTR neck) with a CTC head. About 1.5M parameters, 4.5 MB.
|
| 30 |
+
- **Output vocabulary:** 2,455 classes. Thai is predicted as *grapheme clusters* (a consonant together with its
|
| 31 |
+
upper/lower vowels and tone marks is one class), so stacked marks cannot come out in the wrong order;
|
| 32 |
+
`rec_charset.json` maps each class index straight to its output string.
|
| 33 |
+
- **Input:** BGR line crop resized to height 48 (width kept proportional, max 3200), scaled to [-1, 1].
|
| 34 |
+
- **Training data:** Thai and English text lines. The training data and recipe are not released.
|
| 35 |
+
|
| 36 |
+
### Line-level accuracy (exact match, %)
|
| 37 |
+
|
| 38 |
+
| Split | Lines | Exact | Notes |
|
| 39 |
+
|---|---:|---:|---|
|
| 40 |
+
| printed, documents (`pv_hi_doc`) | 184 | 87.5 | in-house, labels agreed by 3+ of 4 independent readers |
|
| 41 |
+
| printed, photos (`pv_hi_photo`) | 285 | 89.1 | same |
|
| 42 |
+
| SROIE receipts (test) | 1,500 | 48.9 | labels are ALL CAPS, images are mixed case, so case errors count as misses |
|
| 43 |
+
|
| 44 |
+
Measured on the ONNX export with the package's own batching (`Recognizer`, bucket 16, no 180° flip).
|
| 45 |
+
|
| 46 |
+
### Limitations
|
| 47 |
+
|
| 48 |
+
- Handwriting is weak; the model targets printed text.
|
| 49 |
+
- Common remaining confusions are Thai look-alikes: ต/ด, ค/ด, ่/้, ั/ิ, ส/ล, น/บ, ข/บ, ช/ซ.
|
| 50 |
+
- Vertical text and non-Thai/English scripts are not supported.
|
| 51 |
+
|
| 52 |
+
## Conversion
|
| 53 |
+
|
| 54 |
+
`tools/export_tatr.py` and `tools/export_dewarp.py` in the code repository reproduce the TATR and dewarp ONNX
|
| 55 |
+
files. Every export, the recognizer's included, was checked against the original framework on real inputs
|
| 56 |
+
(max abs difference on probabilities ~1e-4 for the recognizer, ~1e-5 for TATR; dewarped pixels within 1 grey
|
| 57 |
+
level on average).
|
NOTICE
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
AksonOCR-Lite
|
| 2 |
+
Copyright 2026 Teetouch Jaknamon
|
| 3 |
+
|
| 4 |
+
This product is licensed under the Apache License, Version 2.0 (see LICENSE).
|
| 5 |
+
|
| 6 |
+
It includes or downloads the following third-party components:
|
| 7 |
+
|
| 8 |
+
1. PaddleX post-processing code (aksonocr_lite/_vendor/paddlex_layout.py, and the
|
| 9 |
+
DB text-detection post-processing in aksonocr_lite/det.py)
|
| 10 |
+
Copyright (c) 2024 PaddlePaddle Authors. Apache License 2.0.
|
| 11 |
+
https://github.com/PaddlePaddle/PaddleX
|
| 12 |
+
|
| 13 |
+
2. Model weights redistributed unmodified in Remtir/AksonOCR-Lite (copies of the official ONNX exports):
|
| 14 |
+
- layout.onnx PP-DocLayoutV3 (PaddlePaddle/PP-DocLayoutV3_onnx) Apache License 2.0
|
| 15 |
+
- det.onnx PP-OCRv6_medium_det (PaddlePaddle/PP-OCRv6_medium_det_onnx) Apache License 2.0
|
| 16 |
+
- doc_ori.onnx PP-LCNet_x1_0_doc_ori (PaddlePaddle/PP-LCNet_x1_0_doc_ori_onnx) Apache License 2.0
|
| 17 |
+
- textline_ori.onnx PP-LCNet_x1_0_textline_ori (PaddlePaddle/PP-LCNet_x1_0_textline_ori_onnx) Apache License 2.0
|
| 18 |
+
Copyright (c) PaddlePaddle Authors.
|
| 19 |
+
|
| 20 |
+
3. Model weights converted to ONNX and redistributed in Remtir/AksonOCR-Lite:
|
| 21 |
+
- tatr.onnx: microsoft/table-transformer-structure-recognition-v1.1-all
|
| 22 |
+
Copyright (c) Microsoft Corporation. MIT License.
|
| 23 |
+
- dewarp_mask.onnx, dewarp_grid.onnx: MonkeyOCRv2 document preprocessor
|
| 24 |
+
(https://github.com/Yuliang-Liu/MonkeyOCRv2). Apache License 2.0.
|
| 25 |
+
- rec.onnx: AksonOCR-Lite Thai recognizer, fine-tuned from PaddlePaddle
|
| 26 |
+
PP-OCRv6_tiny_rec (Apache License 2.0). Released under Apache License 2.0.
|
README.md
ADDED
|
@@ -0,0 +1,253 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- th
|
| 5 |
+
- en
|
| 6 |
+
library_name: onnx
|
| 7 |
+
pipeline_tag: image-to-text
|
| 8 |
+
tags:
|
| 9 |
+
- ocr
|
| 10 |
+
- thai
|
| 11 |
+
- document-parsing
|
| 12 |
+
- text-recognition
|
| 13 |
+
- table-recognition
|
| 14 |
+
- layout-analysis
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
<p align="center">
|
| 18 |
+
<img src="assets/AksonOCR-logo.svg" alt="AksonOCR" width="420">
|
| 19 |
+
</p>
|
| 20 |
+
|
| 21 |
+
<h1 align="center">AksonOCR-Lite</h1>
|
| 22 |
+
|
| 23 |
+
<h2 align="center"><a href="https://aksonocr.com/lite-demo">▶ Try the live demo</a></h2>
|
| 24 |
+
|
| 25 |
+
<p align="center"><b>Code, docs and issues: <a href="https://github.com/Remtir/AksonOCR-Lite">github.com/Remtir/AksonOCR-Lite</a></b></p>
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
Fast, open-source **Thai document OCR** with a grapheme-cluster CTC recognizer. Turns PDFs and images into
|
| 29 |
+
Markdown, plain text or JSON with layout blocks, reading order, text-line boxes and tables. Runs on
|
| 30 |
+
**ONNX Runtime**, on CPU or GPU, with no PyTorch or Paddle install.
|
| 31 |
+
|
| 32 |
+
> **[Try it in your browser →](https://aksonocr.com/lite-demo)** Upload a page image or PDF (up to 7 MB) and see the
|
| 33 |
+
> Markdown, layout and tables. No install.
|
| 34 |
+
|
| 35 |
+
> **81M params, runs on CPU:** ~3× faster per page than 2B-param OCR VLMs on a GPU, and more accurate than VLMs
|
| 36 |
+
> on average across MDPBench, ThaiOCRBench and SEA-DocBench ([benchmarks](BENCHMARKS.md)).
|
| 37 |
+
|
| 38 |
+
```bash
|
| 39 |
+
pip install aksonocr-lite
|
| 40 |
+
aksonocr invoice.pdf > invoice.md
|
| 41 |
+
```
|
| 42 |
+
|
| 43 |
+
- **Thai first:** the text recognizer is trained for Thai (with English), predicting whole grapheme clusters
|
| 44 |
+
so stacked vowels and tone marks come out in the right order.
|
| 45 |
+
- **Document structure:** 25 layout classes (titles, paragraphs, tables, figures, headers/footers, ...) and
|
| 46 |
+
reading order from PP-DocLayoutV3; tables become HTML with row/column spans.
|
| 47 |
+
- **Photos too:** page-orientation correction, upside-down line detection, and optional page dewarping.
|
| 48 |
+
- **Light:** about 330 MB of weights by default (≈620 MB with dewarping), fetched from the Hugging Face Hub on first run.
|
| 49 |
+
|
| 50 |
+
<table><tr><td width="33%"><img src="assets/viz/demo1.gif" alt="Layout blocks and text lines on a Thai government letter" width="100%"></td><td width="33%"><img src="assets/viz/sample2.gif" alt="Layout blocks and text lines on a handwritten page photo" width="100%"></td><td width="33%"><img src="assets/viz/sample1.gif" alt="Layout blocks and text lines on a financial statement" width="100%"></td></tr></table>
|
| 51 |
+
<p align="center"><sub>Layout blocks in reading order, table cells (TATR grid), then detected text lines. Made with <code>tools/visualize_gif.py</code>.</sub></p>
|
| 52 |
+
|
| 53 |
+
## Benchmarks
|
| 54 |
+
|
| 55 |
+
**MDPBench-th 85.6** with `--dewarp` (79.5 default) · **ThaiOCRBench DP 0.640** · **SEA-DocBench-th text NED 0.058**
|
| 56 |
+
· **1.5 s per page on CPU only**.
|
| 57 |
+
|
| 58 |
+
<p align="center">
|
| 59 |
+
<img src="assets/charts/accuracy_light.png" alt="Accuracy of AksonOCR-Lite vs 12 other Thai OCR systems: average of MDPBench-th, ThaiOCRBench Document parsing and SEA-DocBench-th" width="100%">
|
| 60 |
+
</p>
|
| 61 |
+
|
| 62 |
+
<p align="center">
|
| 63 |
+
<img src="assets/charts/speed_light.png" alt="Speed of AksonOCR-Lite vs three VLMs, average over three benchmarks: seconds per page one at a time, pages per second batched" width="100%">
|
| 64 |
+
</p>
|
| 65 |
+
|
| 66 |
+
Averages over MDPBench-th, ThaiOCRBench Document parsing and SEA-DocBench-th.
|
| 67 |
+
**[Full benchmark →](BENCHMARKS.md)** per-benchmark charts, default vs `--dewarp`, speed tables, text recognition
|
| 68 |
+
only.
|
| 69 |
+
|
| 70 |
+
## Requirements
|
| 71 |
+
|
| 72 |
+
| | |
|
| 73 |
+
|---|---|
|
| 74 |
+
| **Python** | 3.10 or newer (CI tests 3.10 and 3.13) |
|
| 75 |
+
| **OS** | Linux, macOS (Apple Silicon). Windows is untested so far. |
|
| 76 |
+
| **Hardware** | Any x86-64 or ARM64 CPU. No GPU needed; NVIDIA GPU optional (see [GPU](#gpu)). |
|
| 77 |
+
| **RAM** | about 1.3 GB peak for one page, about 2.8 GB with `--dewarp` |
|
| 78 |
+
| **Disk** | about 330 MB of model weights, plus 290 MB for `--dewarp`, in the Hugging Face cache (`~/.cache/huggingface`) |
|
| 79 |
+
| **Network** | only on first run, to download the weights. Use [`AKSONOCR_WEIGHTS`](#offline--custom-weights) for fully offline machines. |
|
| 80 |
+
|
| 81 |
+
No PyTorch, Paddle or system packages are needed. `pip` installs everything else: `onnxruntime`, `numpy`,
|
| 82 |
+
`opencv-python-headless`, `pillow`, `pypdfium2` (PDF rendering), `pyclipper`, `shapely` and `huggingface_hub`.
|
| 83 |
+
|
| 84 |
+
For the GPU build (`onnxruntime-gpu`): an NVIDIA driver with CUDA 12 and cuDNN 9, the versions current
|
| 85 |
+
`onnxruntime-gpu` wheels are built against.
|
| 86 |
+
|
| 87 |
+
## Usage
|
| 88 |
+
|
| 89 |
+
### Command line
|
| 90 |
+
|
| 91 |
+
```bash
|
| 92 |
+
aksonocr page.jpg # Markdown to stdout
|
| 93 |
+
aksonocr report.pdf --out out/ # out/report.md (every page of the PDF)
|
| 94 |
+
aksonocr scan.png -f json -o out/ # boxes, blocks, lines, confidences
|
| 95 |
+
aksonocr photo.jpg --dewarp # flatten a photographed / curled page
|
| 96 |
+
aksonocr doc.pdf --device cpu --dpi 200 --max-pages 10
|
| 97 |
+
```
|
| 98 |
+
|
| 99 |
+
### Python
|
| 100 |
+
|
| 101 |
+
```python
|
| 102 |
+
import aksonocr_lite
|
| 103 |
+
|
| 104 |
+
conv = aksonocr_lite.load() # device="auto" | "cpu" | "cuda", dewarp=False, tables=True, ...
|
| 105 |
+
for page in conv.convert("report.pdf"): # path, bytes, PIL.Image or BGR numpy array
|
| 106 |
+
print(page.markdown)
|
| 107 |
+
for block in page.blocks: # reading order
|
| 108 |
+
print(block.label, block.bbox, block.text or block.html)
|
| 109 |
+
for line in page.lines:
|
| 110 |
+
print(line.text, line.confidence, line.poly)
|
| 111 |
+
|
| 112 |
+
lines = conv.ocr(bgr_image) # text lines only, no layout
|
| 113 |
+
```
|
| 114 |
+
|
| 115 |
+
`Page` fields: `markdown`, `text`, `blocks` (`label`, `bbox`, `score`, `order`, `text`, `html`, `line_ids`),
|
| 116 |
+
`lines` (`text`, `confidence`, `bbox`, `poly`), `rotation`, `dewarped`, `width`, `height`, `timings`.
|
| 117 |
+
All coordinates are in the pixel space of the (rotated, dewarped) page image.
|
| 118 |
+
|
| 119 |
+
### GPU
|
| 120 |
+
|
| 121 |
+
The default install runs on CPU, on Linux and macOS (Apple Silicon). For CUDA, swap ONNX Runtime for its GPU build:
|
| 122 |
+
|
| 123 |
+
```bash
|
| 124 |
+
pip uninstall -y onnxruntime && pip install onnxruntime-gpu
|
| 125 |
+
```
|
| 126 |
+
|
| 127 |
+
`device="auto"` then picks CUDA. Inputs have variable shapes, so the CPU path is often just as fast for
|
| 128 |
+
single pages; the GPU pays off mostly with dewarping and large batches.
|
| 129 |
+
|
| 130 |
+
### Offline / custom weights
|
| 131 |
+
|
| 132 |
+
Set `AKSONOCR_WEIGHTS=/path/to/dir` to load models from a local folder instead of the Hub. It must contain
|
| 133 |
+
`layout.onnx det.onnx doc_ori.onnx textline_ori.onnx rec.onnx rec_charset.json tatr.onnx`, plus
|
| 134 |
+
`dewarp_mask.onnx dewarp_grid.onnx` for `--dewarp`.
|
| 135 |
+
|
| 136 |
+
## How it works
|
| 137 |
+
|
| 138 |
+
```
|
| 139 |
+
page image ─┬─ (dewarp: MonkeyOCRv2 preprocessor) optional
|
| 140 |
+
├─ page orientation: PP-LCNet doc_ori 0/90/180/270
|
| 141 |
+
├─ layout + reading order: PP-DocLayoutV3 ┐
|
| 142 |
+
├─ text lines: PP-OCRv6_medium_det (DB) │ run on the same page
|
| 143 |
+
│ └─ line orientation (PP-LCNet textline_ori) │
|
| 144 |
+
│ └─ recognition: AksonOCR Thai CTC (+180° re-read, keeps the more confident reading)
|
| 145 |
+
├─ tables: re-detect text in the crop → TATR v1.1 grid → fill cells with OCR lines → HTML
|
| 146 |
+
└─ merge: each line → the layout block holding its centre (nearest block as fallback),
|
| 147 |
+
blocks in reading order → Markdown
|
| 148 |
+
```
|
| 149 |
+
|
| 150 |
+
### Thai recognizer: grapheme-cluster CTC
|
| 151 |
+
|
| 152 |
+
Thai stacks vowels and tone marks on the consonant, so one glyph can be 2–3 code points (ผู้ = ผ + ู + ้).
|
| 153 |
+
Instead of emitting them one by one, the recognizer predicts the **whole cluster as one token**.
|
| 154 |
+
|
| 155 |
+
<p align="center">
|
| 156 |
+
<img src="assets/viz/cluster_tokens.gif" alt="Step by step: the recognizer's real CTC output on a Thai line, spelled as character tokens (22 symbols, stacked marks as 2–3 symbols in one spot) and as cluster tokens (15 symbols, one per glyph column)" width="100%">
|
| 157 |
+
</p>
|
| 158 |
+
<p align="center"><sub>Real CTC output of the recognizer on a photographed line; made with <code>tools/cluster_gif.py</code>.</sub></p>
|
| 159 |
+
|
| 160 |
+
- **Correct order, always:** marks inside a cluster always come out in Unicode order, never swapped.
|
| 161 |
+
- **No impossible output:** only real clusters are in the vocabulary (2,149 of 2,455 classes).
|
| 162 |
+
- **Shorter sequence:** one token per glyph (15 vs 22 on the line above).
|
| 163 |
+
- **Tiny and fast:** 1.5M parameters, 88.6% exact match at 8 ms per line on one CPU thread
|
| 164 |
+
([comparison](BENCHMARKS.md#text-recognition-only)).
|
| 165 |
+
|
| 166 |
+
Details worth knowing:
|
| 167 |
+
|
| 168 |
+
- Text in tables is read by the Thai recognizer, not by the table model: TATR only provides the grid.
|
| 169 |
+
- The detector often boxes a lone Thai vowel or tone mark; such fragments are dropped before recognition.
|
| 170 |
+
- Lines are batched only with lines of nearly the same width. Padding changes the recognizer's
|
| 171 |
+
attention and cost about 3 points of exact-match accuracy when widths were mixed freely.
|
| 172 |
+
|
| 173 |
+
## Limitations
|
| 174 |
+
|
| 175 |
+
- **Photographed pages are harder than digital ones.** On MDPBench-th the default pipeline scores 89.5 on digital
|
| 176 |
+
pages but 76.2 on photos (curl, shadow, glare, perspective). `--dewarp` brings photos to 84.3 at about twice the
|
| 177 |
+
weights and extra time per page; very dark, folded or blurred photos still lose text.
|
| 178 |
+
- **Handwriting is weak.** The recognizer was trained for printed Thai and English; handwritten notes and
|
| 179 |
+
filled-in forms come out with many errors.
|
| 180 |
+
- **No formula or picture understanding.** Math is read as plain text lines (no LaTeX). Figures and charts
|
| 181 |
+
become `<figure>` blocks holding only the text printed in them, with no description of the picture.
|
| 182 |
+
- **Dense multi-column layouts** such as newspapers often get the reading order or column split wrong.
|
| 183 |
+
- **Thai and English only.** Other scripts and vertical text are not supported.
|
| 184 |
+
|
| 185 |
+
## Models and licenses
|
| 186 |
+
|
| 187 |
+
| Stage | Model | License |
|
| 188 |
+
|---|---|---|
|
| 189 |
+
| Layout + reading order | [PP-DocLayoutV3](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3_onnx) | Apache-2.0 |
|
| 190 |
+
| Text detection | [PP-OCRv6_medium_det](https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_det_onnx) | Apache-2.0 |
|
| 191 |
+
| Page / line orientation | [PP-LCNet_x1_0_doc_ori](https://huggingface.co/PaddlePaddle/PP-LCNet_x1_0_doc_ori_onnx), [textline_ori](https://huggingface.co/PaddlePaddle/PP-LCNet_x1_0_textline_ori_onnx) | Apache-2.0 |
|
| 192 |
+
| Text recognition | AksonOCR Thai recognizer (fine-tuned PP-OCRv6_tiny_rec), [model card](MODEL_CARD.md) | Apache-2.0 |
|
| 193 |
+
| Table structure | [Table Transformer v1.1-all](https://huggingface.co/microsoft/table-transformer-structure-recognition-v1.1-all) (Microsoft) | MIT |
|
| 194 |
+
| Dewarping (optional) | [MonkeyOCRv2](https://github.com/Yuliang-Liu/MonkeyOCRv2) document preprocessor | Apache-2.0 |
|
| 195 |
+
|
| 196 |
+
All weights download from one Hugging Face repo, [Remtir/AksonOCR-Lite](https://huggingface.co/Remtir/AksonOCR-Lite);
|
| 197 |
+
the PaddlePaddle models there are unmodified copies of the official ONNX exports linked above.
|
| 198 |
+
The code is Apache-2.0 ([LICENSE](LICENSE)); third-party attributions are in [NOTICE](NOTICE). Every model
|
| 199 |
+
is permissively licensed, so AksonOCR-Lite can be used commercially. PDFs are rendered with pypdfium2
|
| 200 |
+
(Apache-2.0/BSD), not PyMuPDF (AGPL).
|
| 201 |
+
|
| 202 |
+
## Files in this repository
|
| 203 |
+
|
| 204 |
+
| File | What it is | Origin | License |
|
| 205 |
+
|---|---|---|---|
|
| 206 |
+
| `rec.onnx`, `rec_charset.json` | Thai + English text-line recognizer (ours, [model card](MODEL_CARD.md)) | fine-tuned from PaddlePaddle `PP-OCRv6_tiny_rec` | Apache-2.0 |
|
| 207 |
+
| `layout.onnx` | layout blocks + reading order | [PaddlePaddle/PP-DocLayoutV3_onnx](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3_onnx), unmodified copy | Apache-2.0 |
|
| 208 |
+
| `det.onnx` | text-line detection | [PaddlePaddle/PP-OCRv6_medium_det_onnx](https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_det_onnx), unmodified copy | Apache-2.0 |
|
| 209 |
+
| `doc_ori.onnx`, `textline_ori.onnx` | page / text-line orientation | [PP-LCNet_x1_0_doc_ori_onnx](https://huggingface.co/PaddlePaddle/PP-LCNet_x1_0_doc_ori_onnx), [PP-LCNet_x1_0_textline_ori_onnx](https://huggingface.co/PaddlePaddle/PP-LCNet_x1_0_textline_ori_onnx), unmodified copies | Apache-2.0 |
|
| 210 |
+
| `tatr.onnx` | table structure (rows, columns, headers, spans) | `microsoft/table-transformer-structure-recognition-v1.1-all`, converted | MIT |
|
| 211 |
+
| `dewarp_mask.onnx`, `dewarp_grid.onnx` | page segmentation + dewarping grid (optional) | MonkeyOCRv2 document preprocessor, converted | Apache-2.0 |
|
| 212 |
+
| `config.json` | pipeline manifest (components, source revisions) | ours | Apache-2.0 |
|
| 213 |
+
|
| 214 |
+
The `aksonocr-lite` package downloads these on first use. The four PaddlePaddle files are byte-identical copies of
|
| 215 |
+
the official `inference.onnx` exports, kept here so the whole pipeline loads from one repository.
|
| 216 |
+
|
| 217 |
+
## Development
|
| 218 |
+
|
| 219 |
+
```bash
|
| 220 |
+
pip install -e .[dev]
|
| 221 |
+
pytest -m "not e2e" # unit tests, no weights needed
|
| 222 |
+
pytest # + end-to-end test (downloads weights)
|
| 223 |
+
```
|
| 224 |
+
|
| 225 |
+
`tools/` holds the scripts that produced the ONNX files (they need Paddle/PyTorch and are not part of the
|
| 226 |
+
package): `export_tatr.py`, `export_dewarp.py`.
|
| 227 |
+
|
| 228 |
+
## Acknowledgements
|
| 229 |
+
|
| 230 |
+
AksonOCR-Lite stands on open work from many teams:
|
| 231 |
+
|
| 232 |
+
- [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR) (PaddlePaddle): PP-DocLayoutV3, PP-OCRv6 detection and
|
| 233 |
+
the tiny recognizer the Thai model was fine-tuned from, PP-LCNet orientation models, and the DB / CTC
|
| 234 |
+
post-processing this code ports.
|
| 235 |
+
- [Table Transformer](https://github.com/microsoft/table-transformer) (Microsoft) for table structure.
|
| 236 |
+
- [MonkeyOCRv2](https://github.com/Yuliang-Liu/MonkeyOCRv2) for the page dewarping preprocessor.
|
| 237 |
+
- [ONNX Runtime](https://onnxruntime.ai) and [pypdfium2](https://github.com/pypdfium2-team/pypdfium2), which make
|
| 238 |
+
a framework-free install possible.
|
| 239 |
+
- The benchmark authors: [MDPBench](https://github.com/Yuliang-Liu/MultimodalOCR) (with the OmniDocBench
|
| 240 |
+
evaluator), [ThaiOCRBench](https://github.com/scb-10x/ThaiOCRBench) and
|
| 241 |
+
[SEA-Vision / SEA-DocBench](https://huggingface.co/datasets/xingranzhao/SEA-Vision).
|
| 242 |
+
|
| 243 |
+
## Citation
|
| 244 |
+
|
| 245 |
+
```bibtex
|
| 246 |
+
@software{aksonocr_lite,
|
| 247 |
+
author = {Jaknamon, Teetouch},
|
| 248 |
+
title = {AksonOCR-Lite: Thai Document OCR with Grapheme-Cluster CTC Recognition},
|
| 249 |
+
version = {1.0.0},
|
| 250 |
+
year = {2026},
|
| 251 |
+
url = {https://github.com/Remtir/AksonOCR-Lite}
|
| 252 |
+
}
|
| 253 |
+
```
|
assets/AksonOCR-logo.svg
ADDED
|
|
assets/charts/accuracy_light.png
ADDED
|
Git LFS Details
|
assets/charts/accuracy_mdp_light.png
ADDED
|
Git LFS Details
|
assets/charts/accuracy_sea_light.png
ADDED
|
Git LFS Details
|
assets/charts/accuracy_thai_light.png
ADDED
|
Git LFS Details
|
assets/charts/rec_light.png
ADDED
|
Git LFS Details
|
assets/charts/speed_accuracy_light.png
ADDED
|
Git LFS Details
|
assets/charts/speed_light.png
ADDED
|
Git LFS Details
|
assets/charts/speed_mdp_light.png
ADDED
|
Git LFS Details
|
assets/charts/speed_sea_light.png
ADDED
|
Git LFS Details
|
assets/charts/speed_thai_light.png
ADDED
|
Git LFS Details
|
assets/viz/cluster_tokens.gif
ADDED
|
Git LFS Details
|
assets/viz/demo1.gif
ADDED
|
Git LFS Details
|
assets/viz/sample1.gif
ADDED
|
Git LFS Details
|
assets/viz/sample2.gif
ADDED
|
Git LFS Details
|
bench/RESULTS.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Thai OCR comparison
|
| 2 |
+
|
| 3 |
+
Every system below read the same pages and was scored by the same official scorers. Scripts: `bench/`.
|
| 4 |
+
|
| 5 |
+
## How each system was run
|
| 6 |
+
|
| 7 |
+
- **VLM parsers** (Typhoon OCR V2 2B / 0.8B, Typhoon OCR 1.5 2B, OpenThai2.0-OCR-Markdown, NVIDIA Nemotron 3.5 Parse,
|
| 8 |
+
dots.mocr, Chandra OCR 2): each model's own recipe (prompt, image preprocessing, decoding, token limit, output to
|
| 9 |
+
Markdown) imported unchanged from the public Typhoon OCR V2 evaluation harness
|
| 10 |
+
([srn-source/evaluation_thai_ocr](https://github.com/srn-source/evaluation_thai_ocr), commit `e2512e3`), which
|
| 11 |
+
follows each model card. Served with vLLM 0.33.0 (`vllm serve`, CUDA graphs, 6 API-server processes, up to 256
|
| 12 |
+
requests in flight) on one RTX 5090. No speculative decoding. Typhoon OCR V2 enlarges images whose long side is
|
| 13 |
+
under 1,600 px, as its card requires. No dewarping or page-orientation step was added to any VLM.
|
| 14 |
+
`bench/vlm_serve_bench.py`.
|
| 15 |
+
- **Surya OCR 2.5**: its own package (`surya-ocr` 0.24, full-page mode) against `vllm serve` started with the
|
| 16 |
+
arguments the package uses for its own server; blocks converted to Markdown with Datalab's converter
|
| 17 |
+
(`chandra.output.parse_markdown`). `bench/surya_bench.py`.
|
| 18 |
+
- **Indra OCR**, **OpenThai2.0-OCR-Text**: their packages' defaults on CPU (`bench/cpu_ocr_bench.py`).
|
| 19 |
+
- **AksonOCR-Lite**: default settings (page rotation on, no dewarp), CPU. The `--dewarp` row is the same pipeline
|
| 20 |
+
with MonkeyOCRv2 page flattening first; for this run the dewarp nets ran as one GPU pass over all inputs (saved
|
| 21 |
+
losslessly, identical to `load(dewarp=True)`), the rest on CPU.
|
| 22 |
+
- **Gemini 4 Hydrogen** and **Claude Haiku 5.5** (OpenRouter, reasoning effort low): a generic Markdown prompt;
|
| 23 |
+
ThaiOCRBench with its own item questions (`bench/api_vlm_bench.py`).
|
| 24 |
+
|
| 25 |
+
## Benchmarks
|
| 26 |
+
|
| 27 |
+
| Benchmark | Items | Score | Scorer |
|
| 28 |
+
|---|---:|---|---|
|
| 29 |
+
| MDPBench, Thai | 160 pages (40 digital, 120 photographed) | Overall ↑ | official `pdf_validation.py` + `calculate_scores.py` |
|
| 30 |
+
| ThaiOCRBench, Document parsing | 211 | STEDS-based score ↑ | official `eval_scripts/eval.py` |
|
| 31 |
+
| SEA-DocBench, Thai | 1,506 pages | text NED ↓ (page average), table TEDS ↑ | OmniDocBench end-to-end evaluator (as in SEA-Vision); `bench/seadocbench.py` |
|
| 32 |
+
| pv_hi (internal) | 469 printed text-line crops (184 documents, 285 photos) | exact match % ↑ | `bench/lines_score.py` |
|
| 33 |
+
|
| 34 |
+
Notes:
|
| 35 |
+
- **ThaiOCRBench, two columns.** "question" = the benchmark's own protocol (each item's instruction as the prompt), for
|
| 36 |
+
models that accept a prompt. "page parse" = the model's page-parsing mode, the only mode for systems without a
|
| 37 |
+
prompt. The scorer compares document structure (headings, lists), so converters that write titles without `#`
|
| 38 |
+
(Typhoon OCR V2's doc2json→Markdown) lose most of the score in page-parse mode even with the right text.
|
| 39 |
+
- **SEA-DocBench ground truth** was minimally normalised so the evaluator can read it (formula LaTeX copied to the
|
| 40 |
+
field it reads, category `text` → `text_block`, the 4% of text items with no transcription marked ignore; tables
|
| 41 |
+
carry HTML only, so a prediction written as a LaTeX table is compared with that HTML).
|
| 42 |
+
- **pv_hi** labels are lines on which 3+ of 4 independent readers agreed; one of the readers was an earlier version of
|
| 43 |
+
the AksonOCR recognizer, so this set can favour AksonOCR-Lite. Page parsers read each line crop as a page, which is
|
| 44 |
+
not what they are built for; "recognizer only" is AksonOCR-Lite's recognizer applied to the crop directly.
|
| 45 |
+
- Long generation tails: some VLM outputs repeat until the token limit (counted in `timing.json` per run); they are
|
| 46 |
+
scored as returned.
|
| 47 |
+
|
| 48 |
+
## Results
|
| 49 |
+
|
| 50 |
+
| System | Params | Runs on | MDPBench-th ↑ | digital | photo | ThaiOCRBench DP ↑ (question) | (page parse) | SEA-DocBench-th text NED ↓ | table TEDS ↑ | pv_hi lines exact ↑ (doc / photo) |
|
| 51 |
+
|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|
|
| 52 |
+
| Gemini 4 Hydrogen (reasoning low) | ? | API | 86.5 | 91.3 | 84.9 | 0.576 | 0.372 | 0.0724 | 0.810 | 75.5 / 83.5 |
|
| 53 |
+
| **AksonOCR-Lite** `--dewarp` | 81M + dewarp nets | CPU (dewarp on GPU here), ONNX | 85.6 | 89.5 | 84.3 | – | 0.641 | 0.0595 | 0.757 | 73.9 / 64.6 |
|
| 54 |
+
| Typhoon OCR V2 2B | 2B | vLLM | 84.3 | 85.9 | 83.8 | 0.394 | 0.224 | 0.0626 | 0.888 | 43.5 / 44.9 |
|
| 55 |
+
| **AksonOCR-Lite** (this repo) | 81M | CPU, ONNX | 79.5 | 89.5 | 76.2 | – | 0.640 | 0.0580 | 0.749 | 73.9 / 64.2 (recognizer only: 88.0 / 88.1) |
|
| 56 |
+
| Typhoon OCR V2 0.8B | 0.8B | vLLM | 78.7 | 83.4 | 77.2 | 0.453 | 0.217 | 0.0725 | 0.860 | 33.1 / 35.4 |
|
| 57 |
+
| NVIDIA Nemotron 3.5 Parse | 0.63B | vLLM | 76.5 | 89.3 | 72.2 | – | 0.434 | 0.0494 | 0.865 | 27.2 / 24.6 |
|
| 58 |
+
| Surya OCR 2.5 | 0.69B | vLLM | 75.6 | 85.3 | 72.4 | – | 0.185 | 0.1017 | 0.876 | 17.9 / 14.7 |
|
| 59 |
+
| Chandra OCR 2 | 5B | vLLM | 75.5 | 87.1 | 71.7 | 0.353 | 0.465 | 0.0802 | 0.892 | 12.5 / 21.8 |
|
| 60 |
+
| Typhoon OCR 1.5 2B | 2B | vLLM | 72.9 | 78.0 | 71.2 | 0.252 | 0.292 | 0.0850 | 0.840 | 57.1 / 77.2 |
|
| 61 |
+
| Claude Haiku 5.5 (reasoning low) | ? | API | 70.4 | 89.2 | 64.2 | 0.502 | 0.301 | 0.0684 | 0.818 | 50.0 / 41.4 |
|
| 62 |
+
| OpenThai2.0-OCR-Markdown | 0.63B | vLLM | 70.3 | 83.2 | 66.0 | – | 0.335 | 0.1035 | 0.852 | 32.6 / 30.5 |
|
| 63 |
+
| dots.mocr | 3B | vLLM | 70.3 | 82.3 | 66.3 | 0.257 | 0.219 | 0.0893 | 0.858 | 23.4 / 44.6 |
|
| 64 |
+
| OpenThai2.0-OCR-Text | 1M | CPU, OpenVINO | 58.7 | 72.1 | 54.2 | – | 0.417 | 0.1178 | 0.000 | 32.6 / 59.6 |
|
| 65 |
+
| Indra OCR | ~1M | CPU, ONNX | 36.1 | 75.5 | 22.9 | – | 0.285 | 0.1176 | 0.520 | 4.9 / 14.7 |
|
| 66 |
+
|
| 67 |
+
Generated 2026-11-17 03:26.
|
bench/SPEED.md
ADDED
|
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Speed benchmark
|
| 2 |
+
|
| 3 |
+
Every system is timed on each benchmark's own pages: the 160 Thai MDPBench pages (40 digital, 120 photographed), the
|
| 4 |
+
211 ThaiOCRBench Document parsing images, and every 10th Thai SEA-DocBench page (151 of 1,506). Each page is timed
|
| 5 |
+
end to end, from the image file to the final Markdown (decode, preprocessing, model, post-processing).
|
| 6 |
+
AksonOCR-Lite runs with default settings (page rotation on, tables on, no dewarp). The VLMs run under `vllm serve`
|
| 7 |
+
with the same per-model recipe (prompt, image size, decoding, token limit) as in the accuracy runs. Machine: AMD
|
| 8 |
+
EPYC 7763 (128 logical CPUs), NVIDIA RTX 5090. Raw numbers: [SPEED.json](https://github.com/Remtir/AksonOCR-Lite/blob/main/bench/SPEED.json) (`sets`).
|
| 9 |
+
|
| 10 |
+
"One at a time": one request or page at a time, after warm-up (AksonOCR-Lite: one process with 16 threads; VLMs:
|
| 11 |
+
concurrency 1 on the first 40 pages of the set, 3 of them warm-up). "Batched": AksonOCR-Lite with 16 single-thread
|
| 12 |
+
processes; the VLMs with up to 256 requests in flight (6 API-server processes), over every page of the set.
|
| 13 |
+
|
| 14 |
+
## Average of the three benchmarks
|
| 15 |
+
|
| 16 |
+
| System | Hardware | s / page, one at a time | pages / s, batched | MDPBench-th ↑ | SEA-DocBench-th text NED ↓ |
|
| 17 |
+
|---|---|---:|---:|---:|---:|
|
| 18 |
+
| **AksonOCR-Lite** | CPU only | **1.48** | **1.76** | **79.5** | **0.0580** |
|
| 19 |
+
| **AksonOCR-Lite** | RTX 5090 + CPU | **1.08** | – | **79.5** | **0.0580** |
|
| 20 |
+
| OpenThai2.0-OCR-Markdown | RTX 5090, vLLM | 4.14 | 4.88 | 70.3 | 0.1035 |
|
| 21 |
+
| Typhoon OCR V2 0.8B | RTX 5090, vLLM | 4.17 | 2.14 | 78.7 | 0.0725 |
|
| 22 |
+
| Typhoon OCR V2 2B | RTX 5090, vLLM | 5.11 | 1.23 | 84.3 | 0.0626 |
|
| 23 |
+
|
| 24 |
+
With `--dewarp`, AksonOCR-Lite reaches 85.6 on MDPBench-th, at the cost of one extra dewarp pass per page.
|
| 25 |
+
|
| 26 |
+
## Per benchmark
|
| 27 |
+
|
| 28 |
+
s / page one at a time, pages / s batched (pages that hit the token limit in brackets: they are usually repetition
|
| 29 |
+
loops and cost the full generation budget). Charts: [MDPBench-th](../assets/charts/speed_mdp_light.png) ·
|
| 30 |
+
[ThaiOCRBench DP](../assets/charts/speed_thai_light.png) · [SEA-DocBench-th](../assets/charts/speed_sea_light.png).
|
| 31 |
+
|
| 32 |
+
| System | MDPBench-th (160) | ThaiOCRBench DP (211) | SEA-DocBench-th (151) |
|
| 33 |
+
|---|---|---|---|
|
| 34 |
+
| AksonOCR-Lite, CPU | 1.71 s · 1.57 p/s | 1.31 s · 2.03 p/s | 1.42 s · 1.68 p/s |
|
| 35 |
+
| AksonOCR-Lite, RTX 5090 | 1.15 s | 1.14 s | 0.97 s |
|
| 36 |
+
| OpenThai2.0-OCR-Markdown | 3.91 s · 3.80 p/s | 3.40 s · 7.64 p/s | 5.12 s · 3.21 p/s |
|
| 37 |
+
| Typhoon OCR V2 0.8B | 5.02 s · 0.79 p/s (17) | 3.58 s · 3.09 p/s (1) | 3.90 s · 2.54 p/s (1) |
|
| 38 |
+
| Typhoon OCR V2 2B | 5.29 s · 0.46 p/s (11) | 4.50 s · 1.78 p/s (2) | 5.56 s · 1.46 p/s (2) |
|
| 39 |
+
|
| 40 |
+
MDPBench-th is mostly photographed pages: AksonOCR-Lite spends more on page rotation there, and both Typhoon models
|
| 41 |
+
loop on 11–17 of its pages, which stretches their batched runs.
|
| 42 |
+
|
| 43 |
+
## Earlier mixed set (311 pages)
|
| 44 |
+
|
| 45 |
+
The tables below are the first measurement, on one mixed set: the 160 MDPBench-th pages + the 151 SEA-DocBench-th
|
| 46 |
+
pages above.
|
| 47 |
+
|
| 48 |
+
### One page at a time
|
| 49 |
+
|
| 50 |
+
| System | Hardware | mean s | median s | p90 s | pages |
|
| 51 |
+
|---|---|---:|---:|---:|---:|
|
| 52 |
+
| AksonOCR-Lite | CPU, 1 process × 16 threads | 1.53 | 1.49 | 1.96 | 308 |
|
| 53 |
+
| AksonOCR-Lite | RTX 5090 (`--device cuda`, recognizer on CPU) | 1.08 | 1.07 | 1.48 | 308 |
|
| 54 |
+
| OpenThai2.0-OCR-Markdown | RTX 5090, vLLM, concurrency 1 | 4.01 | 3.91 | 4.97 | 37 |
|
| 55 |
+
| Typhoon OCR V2 0.8B | RTX 5090, vLLM, concurrency 1 | 4.99 | 3.79 | 6.05 | 37 |
|
| 56 |
+
| Typhoon OCR V2 2B | RTX 5090, vLLM, concurrency 1 | 5.34 | 5.18 | 8.19 | 37 |
|
| 57 |
+
|
| 58 |
+
The first 3 pages are warm-up and are not counted. The VLMs are timed on the first 40 pages.
|
| 59 |
+
|
| 60 |
+
AksonOCR-Lite mean time per stage (s):
|
| 61 |
+
|
| 62 |
+
| Stage | CPU | RTX 5090 |
|
| 63 |
+
|---|---:|---:|
|
| 64 |
+
| page rotation | 0.06 | 0.06 |
|
| 65 |
+
| layout | 0.45 | 0.23 |
|
| 66 |
+
| text detection + recognition | 0.81 | 0.60 |
|
| 67 |
+
| tables (structure + cell OCR) | 0.09 | 0.08 |
|
| 68 |
+
|
| 69 |
+
### Batched throughput (all 311 pages)
|
| 70 |
+
|
| 71 |
+
| System | Setup | wall s | pages / s | output tok / s | hit token limit |
|
| 72 |
+
|---|---|---:|---:|---:|---:|
|
| 73 |
+
| AksonOCR-Lite | CPU, 16 processes × 1 thread | 178.9 | 1.74 | – | – |
|
| 74 |
+
| OpenThai2.0-OCR-Markdown | RTX 5090, vLLM, ≤256 in flight | 87.3 | 3.56 | 5,692 | 0 |
|
| 75 |
+
| Typhoon OCR V2 0.8B | RTX 5090, vLLM, ≤256 in flight | 245.0 | 1.27 | 2,622 | 15 |
|
| 76 |
+
| Typhoon OCR V2 2B | RTX 5090, vLLM, ≤256 in flight | 400.2 | 0.78 | 1,224 | 11 |
|
| 77 |
+
|
| 78 |
+
AksonOCR-Lite uses 16 of the 128 logical CPUs and no GPU. Throughput grows with more processes. Pages that hit the
|
| 79 |
+
token limit are usually repetition loops; they cost the full generation budget.
|
| 80 |
+
|
| 81 |
+
## Text recognition only
|
| 82 |
+
|
| 83 |
+
The recognizer alone, on the same 469 pre-cut line crops (internal printed-line set: 184 document + 285 photo
|
| 84 |
+
lines, labels agreed by 3+ of 4 readers), against the other Thai CPU line recognizer, OpenThai2.0-OCR-Text. No
|
| 85 |
+
detection, layout or tables: each recognizer reads one crop file and returns one string, with its own defaults
|
| 86 |
+
(AksonOCR-Lite `aksonocr_lite.rec.Recognizer`; OpenThai `LineOCR(dir, threads=1).recognize([crop])`, its crop
|
| 87 |
+
preprocessing included, OpenVINO). Scored with `bench/lines_score.py`. CPU only, EPYC 7763.
|
| 88 |
+
|
| 89 |
+
| Recognizer | exact % (doc / photo) | exact, ignoring spaces % | character error (NED×100) | ms / line, 1 thread (mean / median / p90) | lines / s, 16 × 1 thread |
|
| 90 |
+
|---|---|---:|---:|---|---:|
|
| 91 |
+
| AksonOCR-Lite (default, with 180° pass) | 88.1 (88.0 / 88.1) | 91.0 | 1.63 | 15.6 / 13.1 / 32.8 | 929 |
|
| 92 |
+
| AksonOCR-Lite, no 180° pass | 88.6 (88.0 / 89.1) | 91.6 | 1.15 | 8.1 / 6.7 / 16.6 | 1,800 |
|
| 93 |
+
| OpenThai2.0-OCR-Text | 38.1 (38.0 / 38.3) | 40.3 | 12.10 | 10.9 / 9.8 / 19.2 | 1,187 |
|
| 94 |
+
|
| 95 |
+
The 180° pass reads every line a second time upside down and keeps the more confident reading: it rescues
|
| 96 |
+
upside-down lines in full pages, costs twice the recognizer time, and on these upright crops changes a few photo
|
| 97 |
+
lines for the worse. The crops were cut by a DB-style detector with some padding; a recognizer trained on tighter
|
| 98 |
+
or looser crops may lose a little here. OpenThai's errors are mostly misread characters
|
| 99 |
+
(e.g. ลักทรัพย์ → ลัคทรัพย์, bullets read as ๑), not formatting. Reproduce:
|
| 100 |
+
`python bench/rec_bench.py --out-dir runs/rec --json bench/SPEED.json`.
|
| 101 |
+
|
| 102 |
+
Page level, for reference: OpenThai2.0-OCR-Text's full pipeline (line detection + recognition + reading order, plain
|
| 103 |
+
text, no layout or tables) in the same CPU setups as above takes 0.33 s per page (OpenVINO; 0.40 s ONNX Runtime) and
|
| 104 |
+
reads 10.2 pages/s with 16 processes (9.9 ONNX Runtime). It does much less work per page than AksonOCR-Lite, whose
|
| 105 |
+
own detection + recognition stage is 0.81 s of its 1.53 s.
|
| 106 |
+
|
| 107 |
+
Reproduce:
|
| 108 |
+
|
| 109 |
+
```bash
|
| 110 |
+
python bench/speed_bench.py --set mdp --out bench/SPEED.json # AksonOCR-Lite on one set: latency, gpu, throughput
|
| 111 |
+
python bench/vlm_speed.py --harness <evaluation_thai_ocr> --model-key typhoon-ocr2-2b --upscale 1600 --sets mdp thai sea --out bench/SPEED.json
|
| 112 |
+
# (typhoon-ocr2-0.8b: --upscale 1600; openthai2.0-ocr-markdown: --upscale 0)
|
| 113 |
+
```
|
config.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_type": "aksonocr-lite",
|
| 3 |
+
"library": "aksonocr-lite",
|
| 4 |
+
"version": "0.1.0",
|
| 5 |
+
"framework": "onnxruntime",
|
| 6 |
+
"languages": [
|
| 7 |
+
"th",
|
| 8 |
+
"en"
|
| 9 |
+
],
|
| 10 |
+
"components": {
|
| 11 |
+
"layout": "layout.onnx",
|
| 12 |
+
"text_detection": "det.onnx",
|
| 13 |
+
"page_orientation": "doc_ori.onnx",
|
| 14 |
+
"textline_orientation": "textline_ori.onnx",
|
| 15 |
+
"text_recognition": "rec.onnx",
|
| 16 |
+
"charset": "rec_charset.json",
|
| 17 |
+
"table_structure": "tatr.onnx",
|
| 18 |
+
"dewarp": [
|
| 19 |
+
"dewarp_mask.onnx",
|
| 20 |
+
"dewarp_grid.onnx"
|
| 21 |
+
]
|
| 22 |
+
},
|
| 23 |
+
"recognizer": {
|
| 24 |
+
"architecture": "SVTR_LCNet",
|
| 25 |
+
"decoder": "grapheme-cluster CTC",
|
| 26 |
+
"input_height": 48
|
| 27 |
+
},
|
| 28 |
+
"sources": {
|
| 29 |
+
"layout.onnx": "PaddlePaddle/PP-DocLayoutV3_onnx@46bbdf1 inference.onnx",
|
| 30 |
+
"det.onnx": "PaddlePaddle/PP-OCRv6_medium_det_onnx@6132380 inference.onnx",
|
| 31 |
+
"doc_ori.onnx": "PaddlePaddle/PP-LCNet_x1_0_doc_ori_onnx@7330ab7 inference.onnx",
|
| 32 |
+
"textline_ori.onnx": "PaddlePaddle/PP-LCNet_x1_0_textline_ori_onnx@7fdcf3c inference.onnx"
|
| 33 |
+
}
|
| 34 |
+
}
|
det.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:eb13b44b25bb36f89528b68720af8a61d9cf381176107f465db1757b65d086e1
|
| 3 |
+
size 62032837
|
dewarp_grid.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e751314fadcf70a0f943a222522c1f1591b2ebb3a0ee8b1ce3e09d029f33ab84
|
| 3 |
+
size 288271378
|
dewarp_mask.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9d59da44df09230ae86f87a260650e024fb35efbb42e43e23be29c3826598755
|
| 3 |
+
size 4615334
|
doc_ori.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:af9a0a4f317ff0709ce752067807f819cb15d883f8ecad89f28df1c6ee2d9c92
|
| 3 |
+
size 6788069
|
layout.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:45bf71750b00739a41fc209f132eb104a4d6b5bb29483c9078164d8b87cf28ba
|
| 3 |
+
size 130502049
|
rec.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e34fd1a1fe5cb5cdefc8696be652cb93b54f62ae8b1cd9a7e41263211ac29564
|
| 3 |
+
size 4504579
|
rec_charset.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"blank": 0, "chars": ["", "!", "\"", "#", "$", "%", "&", "'", "(", ")", "*", "+", ",", "-", ".", "/", "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", ":", ";", "<", "=", ">", "?", "@", "A", "B", "C", "D", "E", "F", "G", "H", "I", "J", "K", "L", "M", "N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z", "[", "\\", "]", "^", "_", "`", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "{", "|", "}", "~", "°", "·", "»", "×", "à", "á", "è", "é", "ì", "í", "ù", "ā", "ē", "ě", "ī", "ō", "ǎ", "ǐ", "ǒ", "ǔ", "Δ", "Ω", "α", "β", "θ", "λ", "π", "ก", "ข", "ฃ", "ค", "ฅ", "ฆ", "ง", "จ", "ฉ", "ช", "ซ", "ฌ", "ญ", "ฎ", "ฏ", "ฐ", "ฑ", "ฒ", "ณ", "ด", "ต", "ถ", "ท", "ธ", "น", "บ", "ป", "ผ", "ฝ", "พ", "ฟ", "ภ", "ม", "ย", "ร", "ฤ", "ล", "ฦ", "ว", "ศ", "ษ", "ส", "ห", "ฬ", "อ", "ฮ", "ฯ", "ะ", "ั", "า", "ำ", "ิ", "ี", "ึ", "ื", "ุ", "ู", "ฺ", "฿", "เ", "แ", "โ", "ใ", "ไ", "ๅ", "ๆ", "็", "่", "้", "๊", "๋", "์", "ํ", "๎", "๏", "๐", "๑", "๒", "๓", "๔", "๕", "๖", "๗", "๘", "๙", "๚", "๛", "–", "—", "•", "…", "※", "⁺", "⁻", "←", "↑", "→", "↓", "↔", "↗", "↘", "↝", "↪", "↳", "↴", "↷", "⇌", "⇒", "⇨", "∅", "∈", "√", "∝", "∧", "∨", "∩", "∴", "≈", "≠", "≡", "≤", "≥", "≫", "⊂", "⊕", "⊖", "⊥", "⎯", "①", "②", "③", "④", "⑤", "⑥", "⑦", "⑧", "Ⓐ", "─", "└", "■", "□", "▪", "▲", "△", "▶", "▷", "▸", "▼", "○", "◎", "●", "◦", "★", "☆", "☐", "☑", "☺", "☼", "♀", "♡", "♥", "⚠", "✅", "✓", "✡", "✧", "✿", "❀", "❌", "❤", "➔", "➜", "➡", "➤", "➥", "⟶", "⤴", "⦿", "⭐", "🌸", "💗", "💜", "📍", "กั", "กั่", "กั้", "กั๊", "กั๋", "กำ", "กิ", "กิ่", "กิ้", "กิ๊", "กิ๋", "กิ์", "กี", "กี่", "กี้", "กี๊", "กี๋", "กึ", "กึ่", "กึ้", "กึ๊", "กึ๋", "กื", "กื่", "กื้", "กื๊", "กื๋", "กุ", "กุ่", "กุ้", "กุ๊", "กุ๋", "กุ์", "กู", "กู่", "กู้", "กู๊", "กู๋", "กฺ", "ก็", "ก่", "ก่ำ", "ก้", "ก้ำ", "ก๊", "ก๊ำ", "ก๋", "ก๋ำ", "ก์", "กํ", "ขั", "ขั่", "ขั้", "ขั๊", "ขั๋", "ขำ", "ขิ", "ขิ่", "ขิ้", "ขิ๊", "ขิ๋", "ขิ์", "ขี", "ขี่", "ขี้", "ขี๊", "ขี๋", "ขึ", "ขึ่", "ขึ้", "ขึ๊", "ขึ๋", "ขื", "ขื่", "ขื้", "ขื๊", "ขื๋", "ขุ", "ขุ่", "ขุ้", "ขุ๊", "ขุ๋", "ขุ์", "ขู", "ขู่", "ขู้", "ขู๊", "ขู๋", "ขฺ", "ข็", "ข่", "ข่ำ", "ข้", "ข้ำ", "ข๊", "ข๊ำ", "ข๋", "ข๋ำ", "ข์", "ขํ", "ฃั", "ฃั่", "ฃั้", "ฃำ", "ฃิ", "ฃิ่", "ฃิ้", "ฃี", "ฃี่", "ฃี้", "ฃึ", "ฃึ่", "ฃึ้", "ฃื", "ฃื่", "ฃื้", "ฃุ", "ฃุ่", "ฃุ้", "ฃู", "ฃู่", "ฃู้", "ฃ็", "ฃ่", "ฃ้", "ฃ๊", "ฃ๋", "ฃ์", "คั", "คั่", "คั้", "คั๊", "คั๋", "คำ", "คิ", "คิ่", "คิ้", "คิ๊", "คิ๋", "คิ์", "คี", "คี่", "คี้", "คี๊", "คี๋", "คึ", "คึ่", "คึ้", "คึ๊", "คึ๋", "คื", "คื่", "คื้", "คื๊", "คื๋", "คุ", "คุ่", "คุ้", "คุ๊", "คุ๋", "คุ์", "คู", "คู่", "คู้", "คู๊", "คู๋", "คฺ", "ค็", "ค่", "ค่ำ", "ค้", "ค้ำ", "ค๊", "ค๊ำ", "ค๋", "ค๋ำ", "ค์", "ฅั", "ฅั่", "ฅั้", "ฅำ", "ฅิ", "ฅิ่", "ฅิ้", "ฅี", "ฅี่", "ฅี้", "ฅึ", "ฅึ่", "ฅึ้", "ฅื", "ฅื่", "ฅื้", "ฅุ", "ฅุ่", "ฅุ้", "ฅู", "ฅู่", "ฅู้", "ฅ็", "ฅ่", "ฅ้", "ฅ๊", "ฅ๋", "ฅ์", "ฆั", "ฆั่", "ฆั้", "ฆั๊", "ฆั๋", "ฆำ", "ฆิ", "ฆิ่", "ฆิ้", "ฆิ๊", "ฆิ๋", "ฆิ์", "ฆี", "ฆี่", "ฆี้", "ฆี๊", "ฆี๋", "ฆึ", "ฆึ่", "ฆึ้", "ฆึ๊", "ฆึ๋", "ฆื", "ฆื่", "ฆื้", "ฆื๊", "ฆื๋", "ฆุ", "ฆุ่", "ฆุ้", "ฆุ๊", "ฆุ๋", "ฆุ์", "ฆู", "ฆู่", "ฆู้", "ฆู๊", "ฆู๋", "ฆฺ", "ฆ็", "ฆ่", "ฆ่ำ", "ฆ้", "ฆ้ำ", "ฆ๊", "ฆ๊ำ", "ฆ๋", "ฆ๋ำ", "ฆ์", "ฆํ", "งั", "งั่", "งั้", "งั๊", "งั๋", "งำ", "งิ", "งิ่", "งิ้", "งิ��", "งิ๋", "งิ์", "งี", "งี่", "งี้", "งี๊", "งี๋", "งึ", "งึ่", "งึ้", "งึ๊", "งึ๋", "งื", "งื่", "งื้", "งื๊", "งื๋", "งุ", "งุ่", "งุ้", "งุ๊", "งุ๋", "งุ์", "งู", "งู่", "งู้", "งู๊", "งู๋", "งฺ", "ง็", "ง่", "ง่ำ", "ง้", "ง้ำ", "ง๊", "ง๊ำ", "ง๋", "ง๋ำ", "ง์", "ง์์", "งํ", "จั", "จั่", "จั้", "จั๊", "จั๋", "จำ", "จิ", "จิ่", "จิ้", "จิ๊", "จิ๋", "จิ์", "จี", "จี่", "จี้", "จี๊", "จี๋", "จึ", "จึ่", "จึ้", "จึ๊", "จึ๋", "จื", "จื่", "จื้", "จื๊", "จื๋", "จุ", "จุ่", "จุ้", "จุ๊", "จุ๋", "จุ์", "จู", "จู่", "จู้", "จู๊", "จู๋", "จฺ", "จ็", "จ่", "จ่ำ", "จ้", "จ้ำ", "จ๊", "จ๊ำ", "จ๋", "จ๋ำ", "จ์", "จํ", "ฉั", "ฉั่", "ฉั้", "ฉั๊", "ฉั๋", "ฉำ", "ฉิ", "ฉิ่", "ฉิ้", "ฉิ๊", "ฉิ๋", "ฉิ์", "ฉี", "ฉี่", "ฉี้", "ฉี๊", "ฉี๋", "ฉึ", "ฉึ่", "ฉึ้", "ฉึ๊", "ฉึ๋", "ฉื", "ฉื่", "ฉื้", "ฉื๊", "ฉื๋", "ฉุ", "ฉุ่", "ฉุ้", "ฉุ๊", "ฉุ๋", "ฉุ์", "ฉู", "ฉู่", "ฉู้", "ฉู๊", "ฉู๋", "ฉฺ", "ฉ็", "ฉ่", "ฉ่ำ", "ฉ้", "ฉ้ำ", "ฉ๊", "ฉ๊ำ", "ฉ๋", "ฉ๋ำ", "ฉ์", "ชั", "ชั่", "ชั้", "ชั๊", "ชั๋", "ชำ", "ชิ", "ชิ่", "ชิ้", "ชิ๊", "ชิ๋", "ชิ์", "ชี", "ชี่", "ชี้", "ชี๊", "ชี๋", "ชึ", "ชึ่", "ชึ้", "ชึ๊", "ชึ๋", "ชื", "ชื่", "ชื้", "ชื๊", "ชื๋", "ชุ", "ชุ่", "ชุ้", "ชุ๊", "ชุ๋", "ชุ์", "ชู", "ชู่", "ชู้", "ชู๊", "ชู๋", "ชฺ", "ช็", "ช่", "ช่ำ", "ช้", "ช้ำ", "ช๊", "ช๊ำ", "ช๋", "ช๋ำ", "ช์", "ซั", "ซั่", "ซั้", "ซั๊", "ซั๋", "ซำ", "ซิ", "ซิ่", "ซิ้", "ซิ๊", "ซิ๋", "ซิ์", "ซี", "ซี่", "ซี้", "ซี๊", "ซี๋", "ซึ", "ซึ่", "ซึ้", "ซึ๊", "ซึ๋", "ซื", "ซื่", "ซื้", "ซื๊", "ซื๋", "ซุ", "ซุ่", "ซุ้", "ซุ๊", "ซุ๋", "ซุ์", "ซู", "ซู่", "ซู้", "ซู๊", "ซู๋", "ซฺ", "ซ็", "ซ่", "ซ่ำ", "ซ้", "ซ้ำ", "ซ๊", "ซ๊ำ", "ซ๋", "ซ๋ำ", "ซ์", "ฌั", "ฌั่", "ฌั้", "ฌั๊", "ฌั๋", "ฌำ", "ฌิ", "ฌิ่", "ฌิ้", "ฌิ๊", "ฌิ๋", "ฌิ์", "ฌี", "ฌี่", "ฌี้", "ฌี๊", "ฌี๋", "ฌึ", "ฌึ่", "ฌึ้", "ฌึ๊", "ฌึ๋", "ฌื", "ฌื่", "ฌื้", "ฌื๊", "ฌื๋", "ฌุ", "ฌุ่", "ฌุ้", "ฌุ๊", "ฌุ๋", "ฌุ์", "ฌู", "ฌู่", "ฌู้", "ฌู๊", "ฌู๋", "ฌฺ", "ฌ็", "ฌ่", "ฌ่ำ", "ฌ้", "ฌ้ำ", "ฌ๊", "ฌ๊ำ", "ฌ๋", "ฌ๋ำ", "ฌ์", "ฌํ", "ญั", "ญั่", "ญั้", "ญั๊", "ญั๋", "ญำ", "ญิ", "ญิ่", "ญิ้", "ญิ๊", "ญิ๋", "ญิ์", "ญี", "ญี่", "ญี้", "ญี๊", "ญี๋", "ญึ", "ญึ่", "ญึ้", "ญึ๊", "ญึ๋", "ญื", "ญื่", "ญื้", "ญื๊", "ญื๋", "ญุ", "ญุ่", "ญุ้", "ญุ๊", "ญุ๋", "ญุ์", "ญู", "ญู่", "ญู้", "ญู๊", "ญู๋", "ญฺ", "ญ็", "ญ่", "ญ่ำ", "ญ้", "ญ้ำ", "ญ๊", "ญ๊ำ", "ญ๋", "ญ๋ำ", "ญ์", "ญํ", "ฎั", "ฎั่", "ฎั้", "ฎั๊", "ฎั๋", "ฎำ", "ฎิ", "ฎิ่", "ฎิ้", "ฎิ๊", "ฎิ๋", "ฎิ์", "ฎี", "ฎี่", "ฎี้", "ฎี๊", "ฎี๋", "ฎึ", "ฎึ่", "ฎึ้", "ฎึ๊", "ฎึ๋", "ฎื", "ฎื่", "ฎื้", "ฎื๊", "ฎื๋", "ฎุ", "ฎุ่", "ฎุ้", "ฎุ๊", "ฎุ๋", "ฎุ์", "ฎู", "ฎู่", "ฎู้", "ฎู๊", "ฎู๋", "ฎฺ", "ฎ็", "ฎ่", "ฎ่ำ", "ฎ้", "ฎ้ำ", "ฎ๊", "ฎ๊ำ", "ฎ๋", "ฎ๋ำ", "ฎ์", "ฎํ", "ฏั", "ฏั่", "ฏั้", "ฏั๊", "ฏั๋", "ฏำ", "ฏิ", "ฏิ่", "ฏิ้", "ฏิ๊", "ฏิ๋", "ฏิ์", "ฏี", "ฏี่", "ฏี้", "ฏี๊", "ฏี๋", "ฏึ", "ฏึ่", "ฏึ้", "ฏึ๊", "ฏึ๋", "ฏื", "ฏื่", "ฏื้", "ฏื๊", "ฏื๋", "ฏุ", "ฏุ่", "ฏุ้", "ฏุ๊", "ฏุ๋", "ฏุ์", "ฏู", "ฏู่", "ฏู้", "ฏู๊", "ฏู๋", "ฏฺ", "ฏ็", "ฏ่", "ฏ่ำ", "ฏ้", "ฏ้ำ", "ฏ๊", "ฏ๊ำ", "ฏ๋", "ฏ๋ำ", "ฏ์", "ฐั", "ฐั่", "ฐั้", "ฐั๊", "ฐั๋", "ฐำ", "ฐิ", "ฐิ่", "ฐิ้", "ฐิ๊", "ฐิ๋", "ฐิ์", "ฐี", "ฐี่", "ฐี้", "ฐี๊", "ฐี๋", "ฐึ", "ฐึ่", "ฐึ้", "ฐึ๊", "ฐึ๋", "ฐื", "ฐื่", "ฐื้", "ฐื๊", "ฐื๋", "ฐุ", "ฐุ่", "ฐุ้", "ฐุ๊", "ฐุ๋", "ฐุ์", "ฐู", "ฐู่", "ฐู้", "ฐู๊", "ฐู๋", "ฐฺ", "ฐ็", "ฐ่", "ฐ่ำ", "ฐ้", "ฐ้ำ", "ฐ๊", "ฐ๊ำ", "ฐ๋", "ฐ๋ำ", "ฐ์", "ฐํ", "ฑั", "ฑั่", "ฑั้", "ฑั๊", "ฑั๋", "ฑำ", "ฑิ", "ฑิ่", "ฑิ้", "ฑิ๊", "ฑิ๋", "ฑิ์", "ฑี", "ฑี่", "ฑี้", "ฑี๊", "ฑี๋", "ฑึ", "ฑึ่", "ฑึ้", "ฑึ๊", "ฑึ๋", "ฑื", "ฑื่", "ฑื้", "ฑื๊", "ฑื๋", "ฑุ", "ฑุ่", "ฑุ้", "ฑุ๊", "ฑุ๋", "ฑุ์", "ฑู", "ฑู่", "ฑู้", "ฑู๊", "ฑู๋", "ฑฺ", "ฑ็", "ฑ่", "ฑ่ำ", "ฑ้", "ฑ้ำ", "ฑ๊", "ฑ๊ำ", "ฑ๋", "ฑ๋ำ", "ฑ์", "ฑํ", "ฒั", "ฒั่", "ฒั้", "ฒั๊", "ฒั๋", "ฒำ", "ฒิ", "ฒิ่", "ฒิ้", "ฒิ๊", "ฒิ๋", "ฒิ์", "ฒี", "ฒี่", "ฒี้", "ฒี๊", "ฒี๋", "ฒึ", "ฒึ่", "ฒึ้", "ฒึ๊", "ฒึ๋", "ฒื", "ฒื่", "ฒื้", "ฒื๊", "ฒื๋", "ฒุ", "ฒุ่", "ฒุ้", "ฒุ๊", "ฒุ๋", "ฒุ์", "ฒู", "ฒู่", "ฒู้", "ฒู๊", "ฒู๋", "ฒฺ", "ฒ็", "ฒ่", "ฒ่ำ", "ฒ้", "ฒ้ำ", "ฒ๊", "ฒ๊ำ", "ฒ๋", "ฒ๋ำ", "ฒ์", "ฒํ", "ณั", "ณั่", "ณั้", "ณั๊", "ณั๋", "ณำ", "ณิ", "ณิ่", "ณิ้", "ณิ๊", "ณิ๋", "ณิ์", "ณี", "ณี่", "ณี้", "ณี๊", "ณี๋", "ณึ", "ณึ่", "ณึ้", "ณึ๊", "ณึ๋", "ณื", "ณื่", "ณื้", "ณื๊", "ณื๋", "ณุ", "ณุ่", "ณุ้", "ณุ๊", "ณุ๋", "ณุ์", "ณู", "ณู่", "ณู้", "ณู๊", "ณู๋", "ณฺ", "ณ็", "ณ่", "ณ่ำ", "ณ้", "ณ้ำ", "ณ๊", "ณ๊ำ", "ณ๋", "ณ๋ำ", "ณ์", "ณํ", "ดั", "ดั่", "ดั้", "ดั๊", "ดั๋", "ดำ", "ดิ", "ดิ่", "ดิ้", "ดิ๊", "ดิ๋", "ดิ์", "ดี", "ดี่", "ดี้", "ดี๊", "ดี๋", "ดึ", "ดึ่", "ดึ้", "ดึ๊", "ดึ๋", "ดื", "ดื่", "ดื้", "ดื๊", "ดื๋", "ดุ", "ดุ่", "ดุ้", "ดุ๊", "ดุ๋", "ดุ์", "ดู", "ดู่", "ดู้", "ดู๊", "ดู๋", "ดฺ", "ด็", "ด่", "ด่ำ", "ด้", "ด้ำ", "ด้่", "ด๊", "ด๊ำ", "ด๋", "ด๋ำ", "ด์", "ดํ", "ตั", "ตั่", "ตั้", "ตั๊", "ตั๋", "ตำ", "ติ", "ติ่", "ติ้", "ติ๊", "ติ๋", "ติ์", "ตี", "ตี่", "ตี้", "ตี๊", "ตี๋", "ตึ", "ตึ่", "ตึ้", "ตึ๊", "ตึ๋", "ตื", "ตื่", "ตื้", "ตื๊", "ตื๋", "ตุ", "ตุ่", "ตุ้", "ตุ๊", "ตุ๋", "ตุ์", "ตู", "ตู่", "ตู้", "ตู๊", "ตู๋", "ตฺ", "ต็", "ต่", "ต่ำ", "ต้", "ต้ำ", "ต๊", "ต๊ำ", "ต๋", "ต๋ำ", "ต์", "ตํ", "ถั", "ถั่", "ถั้", "ถั๊", "ถั๋", "ถำ", "ถิ", "ถิ่", "ถิ้", "ถิ๊", "ถิ๋", "ถิ์", "ถี", "ถี่", "ถี้", "ถี๊", "ถี๋", "ถึ", "ถึ่", "ถึ้", "ถึ๊", "ถึ๋", "ถื", "ถื่", "ถื้", "ถื๊", "ถื๋", "ถุ", "ถุึ", "ถุ่", "ถุ้", "ถุ๊", "ถุ๋", "ถุ์", "ถู", "ถู่", "ถู้", "ถู๊", "ถู๋", "ถฺ", "ถ็", "ถ่", "ถ่ำ", "ถ้", "ถ้ำ", "ถ๊", "ถ๊ำ", "ถ๋", "ถ๋ำ", "ถ์", "ถํ", "ทั", "ทั่", "ทั้", "ทั๊", "ทั๋", "ทำ", "ทิ", "ทิ่", "ทิ้", "ทิ๊", "ทิ๋", "ทิ์", "ที", "ที่", "ที้", "ที๊", "ที๋", "ทึ", "ทึ่", "ทึ้", "ทึ๊", "ทึ๋", "ทื", "ทื่", "ทื้", "ทื๊", "ทื๋", "ทุ", "ทุ่", "ทุ้", "ทุ๊", "ทุ๋", "ทุ์", "ทู", "ทู่", "ทู้", "ทู๊", "ทู๋", "ทฺ", "ท็", "ท่", "ท่ำ", "ท้", "ท้ำ", "ท๊", "ท๊ำ", "ท๋", "ท๋ำ", "ท์", "ธั", "ธั่", "ธั้", "ธั๊", "ธั๋", "ธำ", "ธิ", "ธิ่", "ธิ้", "ธิ๊", "ธิ๋", "ธิ์", "ธี", "ธี่", "ธี้", "ธี๊", "ธี๋", "ธึ", "ธึ่", "ธึ้", "ธึ๊", "ธึ๋", "ธื", "ธื่", "ธื้", "ธื๊", "ธื๋", "ธุ", "ธุ่", "ธุ้", "ธุ๊", "ธุ๋", "ธุ์", "ธู", "ธู่", "ธู้", "ธู๊", "ธู๋", "ธฺ", "ธ็", "ธ่", "ธ่ำ", "ธ้", "ธ้ำ", "ธ๊", "ธ๊ำ", "ธ๋", "ธ๋ำ", "ธ์", "ธํ", "นั", "นั่", "นั้", "นั๊", "นั๋", "นำ", "นิ", "นิ่", "นิ้", "นิ๊", "นิ๋", "นิ์", "นี", "นี่", "นี้", "นี๊", "นี๋", "นึ", "นึ่", "นึ้", "นึ๊", "นึ๋", "นื", "นื่", "นื้", "นื๊", "นื๋", "นุ", "นุ่", "นุ้", "นุ๊", "นุ๋", "นุ์", "นู", "นู่", "นู้", "นู๊", "นู๋", "นฺ", "น็", "น่", "น่ำ", "น้", "น้ำ", "น๊", "น๊ำ", "น๋", "น๋ำ", "น์", "น์์", "นํ", "บั", "บั่", "บั้", "บั๊", "บั๋", "บำ", "บิ", "บิ่", "บิ้", "บิ๊", "บิ๋", "บิ์", "บี", "บี่", "บี้", "บี๊", "บี๋", "บึ", "บึ่", "บึ้", "บึ๊", "บึ๋", "บื", "บื่", "บื้", "บื๊", "บื๋", "บุ", "บุ่", "บุ้", "บุ๊", "บุ๋", "บุ์", "บู", "บู่", "บู้", "บู๊", "บู๋", "บฺ", "บ็", "บ่", "บ่ำ", "บ้", "บ้ำ", "บ๊", "บ๊ำ", "บ๋", "บ๋ำ", "บ์", "ปั", "ปั่", "ปั้", "ปั๊", "ปั๋", "ปำ", "ปิ", "ปิ่", "ปิ้", "ปิ๊", "ปิ๋", "ปิ์", "ปี", "ปี่", "ปี้", "ปี๊", "ปี๋", "ปึ", "ปึ่", "ปึ้", "ปึ๊", "ปึ๋", "ปื", "ปื่", "ปื้", "ปื๊", "ปื๋", "ปุ", "ปุ่", "ปุ้", "ปุ๊", "ปุ๋", "ปุ์", "ปู", "ปู่", "ปู้", "ปู๊", "ปู๋", "ปฺ", "ป็", "ป่", "ป่ำ", "ป้", "ป้ำ", "ป๊", "ป๊ำ", "ป๋", "ป๋ำ", "ป์", "ปํ", "ผั", "ผั่", "ผั้", "ผั๊", "ผั๋", "ผำ", "ผิ", "ผิ่", "ผิ้", "ผิ๊", "ผิ๋", "ผิ์", "ผี", "ผี่", "ผี้", "ผี๊", "ผี๋", "ผึ", "ผึ่", "ผึ้", "ผึ๊", "ผึ๋", "ผื", "ผื่", "ผื้", "ผื๊", "ผื๋", "ผุ", "ผุ่", "ผุ้", "ผุ๊", "ผุ๋", "ผุ์", "ผู", "ผู่", "ผู้", "ผู๊", "ผู๋", "ผฺ", "ผ็", "ผ่", "ผ่ำ", "ผ้", "ผ้ำ", "ผ๊", "ผ๊ำ", "ผ๋", "ผ๋ำ", "ผ์", "ผํ", "ฝั", "ฝั่", "ฝั้", "ฝั๊", "ฝั๋", "ฝำ", "ฝิ", "ฝิ่", "ฝิ้", "ฝิ๊", "ฝิ๋", "ฝิ์", "ฝี", "ฝี่", "ฝี้", "ฝี๊", "ฝี๋", "ฝึ", "ฝึ่", "ฝึ้", "ฝึ๊", "ฝึ๋", "ฝื", "ฝื่", "ฝื้", "ฝื๊", "ฝื๋", "ฝุ", "ฝุ่", "ฝุ้", "ฝุ๊", "ฝุ๋", "ฝุ์", "ฝู", "ฝู่", "ฝู้", "ฝู๊", "ฝู๋", "ฝฺ", "ฝ็", "ฝ่", "ฝ่ำ", "ฝ้", "ฝ้ำ", "ฝ๊", "ฝ๊ำ", "ฝ๋", "ฝ๋ำ", "ฝ์", "ฝํ", "พั", "พั่", "พั้", "พั๊", "พั๋", "พำ", "พิ", "พิ่", "พิ้", "พิ๊", "พิ๋", "พิ์", "พี", "พี่", "พี้", "พี๊", "พี๋", "พึ", "พึ่", "พึ้", "พึ๊", "พึ๋", "พื", "พื่", "พื้", "พื๊", "พื๋", "พุ", "พุ่", "พุ้", "พุ๊", "พุ๋", "พุ์", "พู", "พู่", "พู้", "พู๊", "พู๋", "พฺ", "พ็", "พ่", "พ่ำ", "พ้", "พ้ำ", "พ๊", "พ๊ำ", "พ๋", "พ๋ำ", "พ์", "พํ", "ฟั", "ฟั่", "ฟั้", "ฟั๊", "ฟั๋", "ฟำ", "ฟิ", "ฟิ่", "ฟิ้", "ฟิ๊", "ฟิ๋", "ฟิ์", "ฟี", "ฟี่", "ฟี้", "ฟี๊", "ฟี๋", "ฟึ", "ฟึ่", "ฟึ้", "ฟึ๊", "ฟึ๋", "ฟื", "ฟื่", "ฟื้", "ฟื๊", "ฟื๋", "ฟุ", "ฟุ่", "ฟุ้", "ฟุ๊", "ฟุ๋", "ฟุ์", "ฟู", "ฟู่", "ฟู้", "ฟู๊", "ฟู๋", "ฟฺ", "ฟ็", "ฟ่", "ฟ่ำ", "ฟ้", "ฟ้ำ", "ฟ๊", "ฟ๊ำ", "ฟ๋", "ฟ๋ำ", "ฟ์", "ภั", "ภั่", "ภั้", "ภั๊", "ภั๋", "ภำ", "ภิ", "ภิ่", "ภิ้", "ภิ๊", "ภิ๋", "ภิ์", "ภี", "ภี่", "ภี้", "ภี๊", "ภี๋", "ภึ", "ภึ่", "ภึ้", "ภึ๊", "ภึ๋", "ภื", "ภื่", "ภื้", "ภื๊", "ภื๋", "ภุ", "ภุึ", "ภุ่", "ภุ้", "ภุ๊", "ภุ๋", "ภุ์", "ภู", "ภู่", "ภู้", "ภู๊", "ภู๋", "ภฺ", "ภ็", "ภ่", "ภ่ำ", "ภ้", "ภ้ำ", "ภ๊", "ภ๊ำ", "ภ๋", "ภ๋ำ", "ภ์", "มั", "มั่", "มั้", "มั๊", "มั๋", "มำ", "มิ", "มิ่", "มิ้", "มิ๊", "มิ๋", "มิ์", "มี", "มี่", "มี้", "มี๊", "มี๋", "มึ", "มึ่", "มึ้", "มึ๊", "มึ๋", "มื", "มื่", "มื้", "มื๊", "มื๋", "มุ", "มุ่", "มุ้", "มุ๊", "มุ๋", "มุ์", "มู", "มู่", "มู้", "มู๊", "มู๋", "มฺ", "ม็", "ม่", "ม่ำ", "ม้", "ม้ำ", "ม๊", "ม๊ำ", "ม๋", "ม๋ำ", "ม์", "มํ", "ยั", "ยั่", "ยั้", "ยั๊", "ยั๋", "ยำ", "ยิ", "ยิ่", "ยิ้", "ยิ๊", "ยิ๋", "ยิ์", "ยี", "ยี่", "ยี้", "ยี๊", "ยี๋", "ยึ", "ยึ่", "ยึ้", "ยึ๊", "ยึ๋", "ยื", "ยื่", "ยื้", "ยื๊", "ยื๋", "ยุ", "ยุ่", "ยุ้", "ยุ๊", "ยุ๋", "ยุ์", "ยู", "ยู่", "ยู้", "ยู๊", "ยู๋", "ยฺ", "ย็", "ย่", "ย่ำ", "ย้", "ย้ำ", "ย๊", "ย๊ำ", "ย๋", "ย๋ำ", "ย์", "ยํ", "รั", "รั่", "รั้", "รั๊", "รั๋", "รำ", "ริ", "ริ่", "ริ้", "ริ๊", "ริ๋", "ริ์", "รี", "รี่", "รี้", "รี๊", "รี๋", "รึ", "รึ่", "รึ้", "รึ๊", "รึ๋", "รื", "รื่", "รื้", "รื๊", "รื๋", "รุ", "รุ่", "รุ้", "รุ๊", "รุ๋", "รุ์", "รู", "รู่", "รู้", "รู๊", "รู๋", "รฺ", "ร็", "ร่", "ร่ำ", "ร้", "ร้ำ", "ร๊", "ร๊ำ", "ร๋", "ร๋ำ", "ร์", "รํ", "ลั", "ลั่", "ลั้", "ลั๊", "ลั๋", "ลำ", "ลิ", "ลิ่", "ลิ้", "ลิ๊", "ลิ๋", "ลิ์", "ลี", "ลี่", "ลี้", "ลี๊", "ลี๋", "ลึ", "ลึ่", "ลึ้", "ลึ๊", "ลึ๋", "ลื", "ลื่", "ลื้", "ลื๊", "ลื๋", "ลุ", "ลุ่", "ลุ้", "ลุ๊", "ลุ๋", "ลุ์", "ลู", "ลู่", "ลู้", "ลู๊", "ลู๋", "ลฺ", "ล็", "ล่", "ล่ำ", "ล้", "ล้ำ", "ล๊", "ล๊ำ", "ล๋", "ล๋ำ", "ล์", "ลํ", "วั", "วั่", "วั้", "วั๊", "วั๋", "วำ", "วิ", "วิ่", "วิ้", "วิ๊", "วิ๋", "วิ์", "วี", "วี่", "วี้", "วี๊", "วี๋", "วึ", "วึ่", "วึ้", "วึ๊", "วึ๋", "วื", "วื่", "วื้", "วื๊", "วื๋", "วุ", "วุ่", "วุ้", "วุ๊", "วุ๋", "วุ์", "วู", "วู่", "วู้", "วู๊", "วู๋", "วฺ", "ว็", "ว่", "ว่ำ", "ว้", "ว้ำ", "ว๊", "ว๊ำ", "ว๋", "ว๋ำ", "ว์", "วํ", "ศั", "ศั่", "ศั้", "ศั๊", "ศั๋", "ศำ", "ศิ", "ศิ่", "ศิ้", "ศิ๊", "ศิ๋", "ศิ์", "ศี", "ศี่", "ศี้", "ศี๊", "ศี๋", "ศึ", "ศึ่", "ศึ้", "ศึ๊", "ศึ๋", "ศื", "ศื่", "ศื้", "ศื๊", "ศื๋", "ศุ", "ศุ่", "ศุ้", "ศุ๊", "ศุ๋", "ศุ์", "ศู", "ศู่", "ศู้", "ศู๊", "ศู๋", "ศฺ", "ศ็", "ศ่", "ศ่ำ", "ศ้", "ศ้ำ", "ศ๊", "ศ๊ำ", "ศ๋", "ศ๋ำ", "ศ์", "ษั", "ษั่", "ษั้", "ษั๊", "ษั๋", "ษำ", "ษิ", "ษิ่", "ษิ้", "ษิ๊", "ษิ๋", "ษิ์", "ษี", "ษี่", "ษี้", "ษี๊", "ษี๋", "ษึ", "ษึ่", "ษึ้", "ษึ๊", "ษึ๋", "ษื", "ษื่", "ษื้", "ษื๊", "ษื๋", "ษุ", "ษุ่", "ษุ้", "ษุ๊", "ษุ๋", "ษุ์", "ษู", "ษู่", "ษู้", "ษู๊", "ษู๋", "ษฺ", "ษ็", "ษ่", "ษ่ำ", "ษ้", "ษ้ำ", "ษ๊", "ษ๊ำ", "ษ๋", "ษ๋ำ", "ษ์", "ษํ", "สั", "สั่", "สั้", "สั๊", "สั๋", "สำ", "สิ", "สิ่", "สิ้", "สิ๊", "สิ๋", "สิ์", "สี", "สี่", "สี้", "สี๊", "สี๋", "สึ", "สึ่", "สึ้", "สึ๊", "สึ๋", "สื", "สื่", "สื้", "สื๊", "สื๋", "สุ", "สุ่", "สุ้", "สุ๊", "สุ๋", "สุ์", "สู", "สู่", "สู้", "สู๊", "สู๋", "สฺ", "ส็", "ส่", "ส่ำ", "ส้", "ส้ำ", "ส๊", "ส๊ำ", "ส๋", "ส๋ำ", "ส์", "สํ", "หั", "หั่", "หั้", "หั๊", "หั๋", "หำ", "หิ", "หิ่", "หิ้", "หิ๊", "หิ๋", "หิ์", "หี", "หี่", "หี้", "หี๊", "หี๋", "หึ", "หึ่", "หึ้", "หึ๊", "หึ๋", "หื", "หื่", "หื้", "หื๊", "หื๋", "หุ", "หุ่", "หุ้", "หุ๊", "หุ๋", "หุ์", "หู", "หู่", "หู้", "หู๊", "หู๋", "หฺ", "ห็", "ห่", "ห่ำ", "ห้", "ห้ำ", "ห๊", "ห๊ำ", "ห๋", "ห๋ำ", "ห์", "หํ", "ฬั", "ฬั่", "ฬั้", "ฬั๊", "ฬั๋", "ฬำ", "ฬิ", "ฬิ่", "ฬิ้", "ฬิ๊", "ฬิ๋", "ฬิ์", "ฬี", "ฬี่", "ฬี้", "ฬี๊", "ฬี๋", "ฬึ", "ฬึ่", "ฬึ้", "ฬึ๊", "ฬึ๋", "ฬื", "ฬื่", "ฬื้", "ฬื๊", "ฬื๋", "ฬุ", "ฬุ่", "ฬุ้", "ฬุ๊", "ฬุ๋", "ฬุ์", "ฬู", "ฬู่", "ฬู้", "ฬู๊", "ฬู๋", "ฬฺ", "ฬ็", "ฬ่", "ฬ่ำ", "ฬ้", "ฬ้ำ", "ฬ๊", "ฬ๊ำ", "ฬ๋", "ฬ๋ำ", "ฬ์", "อั", "อั่", "อั้", "อั๊", "อั๋", "อำ", "อิ", "อิ่", "อิ้", "อิ๊", "อิ๋", "อิ์", "อี", "อี่", "อี้", "อี๊", "อี๋", "อึ", "อึ่", "อึ้", "อึ๊", "อึ๋", "อื", "อื่", "อื้", "อื๊", "อื๋", "อุ", "อุ่", "อุ้", "อุ๊", "อุ๋", "อุ์", "อู", "อู่", "อู้", "อู๊", "อู๋", "อฺ", "อ็", "อ่", "อ่ำ", "อ้", "อ้ำ", "อ๊", "อ๊ำ", "อ๋", "อ๋ำ", "อ์", "อํ", "ฮั", "ฮั่", "ฮั้", "ฮั๊", "ฮั๋", "ฮำ", "ฮิ", "ฮิ่", "ฮิ้", "ฮิ๊", "ฮิ๋", "ฮิ์", "ฮี", "ฮี่", "ฮี้", "ฮี๊", "ฮี๋", "ฮึ", "ฮึ่", "ฮึ้", "ฮึ๊", "ฮึ๋", "ฮื", "ฮื่", "ฮื้", "ฮื๊", "ฮื๋", "ฮุ", "ฮุ่", "ฮุ้", "ฮุ๊", "ฮุ๋", "ฮุ์", "ฮู", "ฮู่", "ฮู้", "ฮู๊", "ฮู๋", "ฮฺ", "ฮ็", "ฮ่", "ฮ่ำ", "ฮ้", "ฮ้ำ", "ฮ๊", "ฮ๊ำ", "ฮ๋", "ฮ๋ำ", "ฮ์", " "]}
|
tatr.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d8882704aa5f96a773dd3f5caea929e9270dbb3885254af022242983ade27e30
|
| 3 |
+
size 115816384
|
textline_ori.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:38aa97cd4be591e0ad304e659f07ba30d946f27a63315433f6659c69c8778345
|
| 3 |
+
size 6777816
|