TeetouchQQ commited on
Commit
64e20ae
·
0 Parent(s):

AksonOCR-Lite 1.0.0 models

Browse files
.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
+ ![Accuracy, average of three benchmarks](assets/charts/accuracy_light.png)
43
+
44
+ ### MDPBench-th
45
+
46
+ ![MDPBench-th, overall](assets/charts/accuracy_mdp_light.png)
47
+
48
+ ### ThaiOCRBench, Document parsing
49
+
50
+ ![ThaiOCRBench Document parsing](assets/charts/accuracy_thai_light.png)
51
+
52
+ ### SEA-DocBench-th
53
+
54
+ ![SEA-DocBench-th text error](assets/charts/accuracy_sea_light.png)
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
+ ![Speed, average of three benchmarks](assets/charts/speed_light.png)
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
+ ![Speed on MDPBench-th pages](assets/charts/speed_mdp_light.png)
80
+
81
+ ### ThaiOCRBench, Document parsing
82
+
83
+ ![Speed on ThaiOCRBench pages](assets/charts/speed_thai_light.png)
84
+
85
+ ### SEA-DocBench-th
86
+
87
+ ![Speed on SEA-DocBench-th pages](assets/charts/speed_sea_light.png)
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
+ ![Text recognition only](assets/charts/rec_light.png)
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

  • SHA256: 35b9be448614eac276e5c9ec6ac80017d8bf133e3b54b5e5e24e35f960c4fcb4
  • Pointer size: 131 Bytes
  • Size of remote file: 293 kB
assets/charts/accuracy_mdp_light.png ADDED

Git LFS Details

  • SHA256: 26632cdbbb4bdb5ab7315d9a9fe39036cfa0b8737f7ca05ccfc11201cefa3bc9
  • Pointer size: 131 Bytes
  • Size of remote file: 270 kB
assets/charts/accuracy_sea_light.png ADDED

Git LFS Details

  • SHA256: b061cd8dc04fe79e793ccc33b4b5cad6a89e6c8b5a0f899ba8480ac0b08eb49d
  • Pointer size: 131 Bytes
  • Size of remote file: 259 kB
assets/charts/accuracy_thai_light.png ADDED

Git LFS Details

  • SHA256: 8cabbebf576056a5c7bc3c49b094ceee64f27ee65c6fe35754780ce6940d4879
  • Pointer size: 131 Bytes
  • Size of remote file: 270 kB
assets/charts/rec_light.png ADDED

Git LFS Details

  • SHA256: 6a87a2cd66aa3339223d32c5c02c37d275751799aa821c79edc99f31b503cab7
  • Pointer size: 131 Bytes
  • Size of remote file: 205 kB
assets/charts/speed_accuracy_light.png ADDED

Git LFS Details

  • SHA256: 24d456efd3164e74e4839ee77745b14dcdc52791e48da2005fc715c491d576f8
  • Pointer size: 131 Bytes
  • Size of remote file: 177 kB
assets/charts/speed_light.png ADDED

Git LFS Details

  • SHA256: 7b2ca6ded8f62a118e5f155727e2fececeb5c4b48023042c6862bfcfd1642bb6
  • Pointer size: 131 Bytes
  • Size of remote file: 294 kB
assets/charts/speed_mdp_light.png ADDED

Git LFS Details

  • SHA256: 2aeb57d022221f980270d66bd4f9e74633a75d85cb530bbb5c747315103b0896
  • Pointer size: 131 Bytes
  • Size of remote file: 265 kB
assets/charts/speed_sea_light.png ADDED

Git LFS Details

  • SHA256: e6ef421721d7c9ca3547a6cc97efccb571c3fee516c8e28d7ddf2575b8a28642
  • Pointer size: 131 Bytes
  • Size of remote file: 278 kB
assets/charts/speed_thai_light.png ADDED

Git LFS Details

  • SHA256: 835e1d924d846441ab7985afaf706a61bd5493ce1ee4c1372956069916e9df45
  • Pointer size: 131 Bytes
  • Size of remote file: 275 kB
assets/viz/cluster_tokens.gif ADDED

Git LFS Details

  • SHA256: c5bd884cf24f93a7f136b488d08ab39c0047bdfdf2d66a6aca73502d297eace3
  • Pointer size: 131 Bytes
  • Size of remote file: 945 kB
assets/viz/demo1.gif ADDED

Git LFS Details

  • SHA256: f1438d564e7f25ddb64b1a830f1fe53f3790d5c06ec299e9826c87dbe38dee32
  • Pointer size: 131 Bytes
  • Size of remote file: 309 kB
assets/viz/sample1.gif ADDED

Git LFS Details

  • SHA256: da96dbd04ca28f61979fcc8fdb65dd20f1c2494f215bd3ff487185c365ab8a36
  • Pointer size: 131 Bytes
  • Size of remote file: 194 kB
assets/viz/sample2.gif ADDED

Git LFS Details

  • SHA256: cdc4b80e6d6c2ca93bfde67a7ba84346950ee051622f890f8b32e888f4911190
  • Pointer size: 131 Bytes
  • Size of remote file: 621 kB
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