Datasets:

Formats:
json
ArXiv:
wdx360 commited on
Commit
4273e37
·
1 Parent(s): 6a6fb69

update readme

Browse files
Files changed (3) hide show
  1. README.md +4 -4
  2. README_zh-CN.md +577 -0
  3. assets/InduOCRBench_overview.png +3 -0
README.md CHANGED
@@ -2,8 +2,9 @@
2
 
3
  <div align="center">
4
  English | <a href="./README_zh-CN.md">简体中文</a>
 
5
 
6
- [\[📜 arXiv\]]() | [[Dataset (🤗Hugging Face)]]()
7
 
8
  ---
9
  ## News
@@ -11,11 +12,9 @@ English | <a href="./README_zh-CN.md">简体中文</a>
11
  - **[2026-04]** InduOCRBench paper accepted to ACL 2026 Industry Track. Dataset released.
12
 
13
  ---
14
- <p align="center">
15
 
16
  ![](assets/InduOCRBench_overview.png)
17
 
18
- </p>
19
 
20
  ---
21
 
@@ -62,6 +61,7 @@ English | <a href="./README_zh-CN.md">简体中文</a>
62
  - VisualStyle
63
  - Watermark
64
  - CrosspageTable
 
65
  ---
66
 
67
  ## 📂 Dataset Structure
@@ -367,7 +367,7 @@ If you use InduOCRBench in your research, please consider citing:
367
  title={When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation},
368
  author={Lin Sun and Wangdexian and Jingang Huang and Linglin Zhang and Change Jia and Zhengwei Cheng and Xiangzheng Zhang},
369
  year={2026},
370
- eprint={},
371
  archivePrefix={arXiv},
372
  primaryClass={cs.CV},
373
  url={https://github.com/Qihoo360/InduOCRBench},
 
2
 
3
  <div align="center">
4
  English | <a href="./README_zh-CN.md">简体中文</a>
5
+ </div>
6
 
7
+ [\[📜 arXiv\]](https://arxiv.org/abs/2605.00911) | [[Dataset (🤗Hugging Face)]](https://huggingface.co/datasets/qihoo360/InduOCRBench)
8
 
9
  ---
10
  ## News
 
12
  - **[2026-04]** InduOCRBench paper accepted to ACL 2026 Industry Track. Dataset released.
13
 
14
  ---
 
15
 
16
  ![](assets/InduOCRBench_overview.png)
17
 
 
18
 
19
  ---
20
 
 
61
  - VisualStyle
62
  - Watermark
63
  - CrosspageTable
64
+
65
  ---
66
 
67
  ## 📂 Dataset Structure
 
367
  title={When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation},
368
  author={Lin Sun and Wangdexian and Jingang Huang and Linglin Zhang and Change Jia and Zhengwei Cheng and Xiangzheng Zhang},
369
  year={2026},
370
+ eprint={2605.00911},
371
  archivePrefix={arXiv},
372
  primaryClass={cs.CV},
373
  url={https://github.com/Qihoo360/InduOCRBench},
README_zh-CN.md ADDED
@@ -0,0 +1,577 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <h1 align="center">InduOCRBench</h1>
2
+
3
+ <div align="center">
4
+ <a href="./README.md">English</a> | 简体中文
5
+ </div>
6
+
7
+ [[📜 arXiv]](https://arxiv.org/abs/2605.00911) | [[Dataset (🤗Hugging Face)]](https://huggingface.co/datasets/qihoo360/InduOCRBench)
8
+
9
+ ---
10
+
11
+ ## News
12
+
13
+ - **[2026-04]** InduOCRBench 论文被 ACL 2026 Industry Track 接收,数据集正式发布。
14
+
15
+ ---
16
+ <p align="center">
17
+
18
+ ![](assets/InduOCRBench_overview.png)
19
+
20
+ </p>
21
+
22
+ ---
23
+
24
+ ## 📖 项目简介
25
+
26
+ **InduOCRBench** 是一个面向工业级 RAG 系统的 OCR Benchmark,覆盖真实企业场景中常见的 11 类高挑战文档类型。
27
+
28
+ 该 Benchmark 聚焦于传统字符级 OCR 指标与真实下游 RAG 效果之间的鸿沟,从:
29
+
30
+ - OCR 识别保真度(OCR Fidelity)
31
+ - 端到端 RAG 检索与问答效果(RAG Impact)
32
+
33
+ 两个维度系统评估 OCR 的真实鲁棒性。
34
+
35
+ ---
36
+
37
+ ## ✨ 核心特点
38
+
39
+ - **真实工业场景**
40
+ - 数据来源于 12 个行业、1 万份真实文档。
41
+
42
+ - **大规模高多样性**
43
+ - 包含 **570** 份 PDF 文档、共 **3,402** 页。
44
+ - 覆盖 **11 类 OCR 挑战场景 + 1 类 Normal 场景**。
45
+
46
+ - **高质量标注**
47
+ - 使用细粒度 Hybrid Markdown 标注:
48
+ - Markdown
49
+ - HTML 表格
50
+ - LaTeX 公式
51
+ - Style Tags
52
+ - 采用 3-stage human-in-the-loop 质检流程,标注准确率达到 98%。
53
+
54
+ - **双评测体系**
55
+ - OCR Fidelity(字符 / 结构级评测)
56
+ - RAG Impact(端到端检索与生成评测)
57
+
58
+ ---
59
+
60
+ ## 🔍 关键发现
61
+
62
+ - 在标准 Benchmark(如 OmniDocBench)上接近满分的模型,在 InduOCRBench 上出现明显性能下降:
63
+ - PP-StructureV3 ↓ 26.4 pts
64
+ - PaddleOCR-VL ↓ 14.7 pts
65
+
66
+ - **高 OCR 准确率并不意味着高 RAG 效果。**
67
+ - `VisualStyle` 文档 OCR Accuracy 达到 82.9%
68
+ - 但 RAG Accuracy 仅有 52.8%
69
+ - 两者存在 30.1 pts 的巨大差距。
70
+
71
+ - OCR 引起的信息缺失,是所有 OCR-first RAG 架构中的稳定上游瓶颈。
72
+
73
+ ---
74
+
75
+ ## 📦 Benchmark 包含两个评测任务
76
+
77
+ ### 1. OCR Fidelity Evaluation
78
+
79
+ 基于 Ground-truth Markdown,对 OCR 输出进行字符级与结构级评测。
80
+
81
+ 对应目录:
82
+
83
+ ```text
84
+ ocr_data/
85
+ ```
86
+
87
+ ---
88
+
89
+ ### 2. RAG Impact Evaluation
90
+
91
+ 评估 OCR 质量对端到端:
92
+
93
+ - Retrieval
94
+ - Generation
95
+ - QA Accuracy
96
+
97
+ 的影响。
98
+
99
+ 对应目录:
100
+
101
+ ```text
102
+ RAG_eval/
103
+ ```
104
+
105
+ ---
106
+
107
+ # 📊 数据集统计
108
+
109
+ | 统计项 | 数值 |
110
+ |---|---|
111
+ | 文档数量 | 570 |
112
+ | 页面数量 | 3,402 |
113
+ | 文档类型 | 11 类挑战场景 + 1 类 Normal |
114
+ | QA 数量(RAG) | 2071 |
115
+ | 标注格式 | Hybrid Markdown |
116
+
117
+ ---
118
+
119
+ ## 11 类 OCR 挑战文档类型
120
+
121
+ - ComplexBackground
122
+ - HighPixel
123
+ - UltraLong
124
+ - MultiColumn
125
+ - UltraWide
126
+ - HistoryBooks
127
+ - Handwriting
128
+ - MultiFont
129
+ - VisualStyle
130
+ - Watermark
131
+ - CrosspageTable
132
+
133
+ ---
134
+
135
+ # 📂 数据集结构
136
+
137
+ ```text
138
+ InduOCRBench/
139
+ ├── ocr_data/
140
+ │ ├── pdf.zip # 原始 PDF 文档(570份,3402页)
141
+ │ ├── md.zip # 【推荐】OCR 评测 Ground Truth
142
+ │ └── md_original.zip # 保留完整视觉样式信息的高保真标注
143
+
144
+ ├── RAG_eval/
145
+ │ ├── QA_pairs.jsonl # RAG 评测 QA 数据
146
+ │ └── doc_md/ # QA 对应的 Ground Truth Markdown
147
+
148
+ ├── README.md
149
+ └── README_zh-CN.md
150
+ ```
151
+
152
+ ---
153
+
154
+ ## 各类 Markdown 文件说明
155
+
156
+ ### md_original
157
+
158
+ 高保真 Markdown 标注版本,保留:
159
+
160
+ - 字体
161
+ - 颜色
162
+ - 对齐方式
163
+ - 布局
164
+ - 视觉 style tags
165
+
166
+ 适用于:
167
+
168
+ - 文档重建
169
+ - 高保真 OCR
170
+ - 文档视觉理解研究
171
+
172
+ ---
173
+
174
+ ### md
175
+
176
+ 移除了视觉样式,仅保留文本内容。
177
+
178
+ 该版本作为:
179
+
180
+ > OCR Fidelity Evaluation 标准 Ground Truth
181
+
182
+ 用于保证评测公平性。
183
+
184
+ ---
185
+
186
+ ### doc_md
187
+
188
+ 用于 RAG 构建的 Hybrid Markdown。
189
+
190
+ 其中:
191
+
192
+ - `VisualStyle` 文档保留 style 信息
193
+ - 其他类型移除 style 信息
194
+
195
+ 该版本作为:
196
+
197
+ > RAG Indexing 与 QA Evaluation 的标准 Ground Truth
198
+
199
+ ---
200
+
201
+ # 🚀 OCR Evaluation
202
+
203
+ 本 Benchmark 使用 **OmniDocBench** 中的 `md2md` 评测方法。
204
+
205
+ 详情参考:
206
+
207
+ https://github.com/opendatalab/OmniDocBench/tree/main
208
+
209
+ ---
210
+
211
+ ## Evaluation Result
212
+
213
+ <div align="center">
214
+
215
+ Table 1. OCR fidelity evaluation on InduOCRBench using md2md metrics
216
+
217
+ </div>
218
+
219
+ <table style="width:100%; border-collapse: collapse;">
220
+ <thead>
221
+ <tr>
222
+ <th>Model Type</th>
223
+ <th>Methods</th>
224
+ <th>Size</th>
225
+ <th>Overall&#x2191;</th>
226
+ <th>Text<sup>EDS</sup>&#x2191;</th>
227
+ <th>Formula<sup>CDM</sup>&#x2191;</th>
228
+ <th>Table<sup>TEDS</sup>&#x2191;</th>
229
+ <th>Table<sup>TEDS-S</sup>&#x2191;</th>
230
+ <th>Read Order<sup>EDS</sup>&#x2191;</th>
231
+ </tr>
232
+ </thead>
233
+ <tbody>
234
+ <tr>
235
+ <td rowspan="9"><strong>Specialized</strong><br><strong>VLMs</strong></td>
236
+ <td>PaddleOCR-VL-1.5</td>
237
+ <td>0.9B</td>
238
+ <td><strong>79.01</strong></td>
239
+ <td><strong>88.33</strong></td>
240
+ <td>75.3</td>
241
+ <td><strong>73.41</strong></td>
242
+ <td><strong>77.27</strong></td>
243
+ <td>85.3</td>
244
+ </tr>
245
+ <tr>
246
+ <td>PaddleOCR-VL</td>
247
+ <td>0.9B</td>
248
+ <td><ins>78.24</ins></td>
249
+ <td><ins>88.1</ins></td>
250
+ <td>74.6</td>
251
+ <td><ins>72.03</ins></td>
252
+ <td>75.87</td>
253
+ <td>85.6</td>
254
+ </tr>
255
+ <tr>
256
+ <td>Logics-Parsing-v2</td>
257
+ <td>4B</td>
258
+ <td>75.71</td>
259
+ <td>84.94</td>
260
+ <td>72.3</td>
261
+ <td>69.90</td>
262
+ <td>76.17</td>
263
+ <td><strong>88.9</strong></td>
264
+ </tr>
265
+ <tr>
266
+ <td>MinerU2.5-Pro</td>
267
+ <td>1.2B</td>
268
+ <td>74.47</td>
269
+ <td>81.63</td>
270
+ <td><ins>75.8</ins></td>
271
+ <td>65.99</td>
272
+ <td>70.46</td>
273
+ <td>79.1</td>
274
+ </tr>
275
+ <tr>
276
+ <td>FireRed-OCR</td>
277
+ <td>2B</td>
278
+ <td>74.09</td>
279
+ <td>87.9</td>
280
+ <td>72.4</td>
281
+ <td>61.98</td>
282
+ <td>66.45</td>
283
+ <td><ins>85.8</ins></td>
284
+ </tr>
285
+ <tr>
286
+ <td>MinerU2.5</td>
287
+ <td>1.2B</td>
288
+ <td>72.50</td>
289
+ <td>81.8</td>
290
+ <td>75.4</td>
291
+ <td>60.31</td>
292
+ <td>63.10</td>
293
+ <td>84.4</td>
294
+ </tr>
295
+ <tr>
296
+ <td>GLM-OCR</td>
297
+ <td>0.9B</td>
298
+ <td>68.64</td>
299
+ <td>63.18</td>
300
+ <td>72.1</td>
301
+ <td>70.64</td>
302
+ <td><ins>76.72</ins></td>
303
+ <td>77.2</td>
304
+ </tr>
305
+ <tr>
306
+ <td>hunyuan-ocr</td>
307
+ <td>0.9B</td>
308
+ <td>68.08</td>
309
+ <td>86.1</td>
310
+ <td>65.6</td>
311
+ <td>52.53</td>
312
+ <td>58.34</td>
313
+ <td>85.7</td>
314
+ </tr>
315
+ <tr>
316
+ <td>deepseek-ocr</td>
317
+ <td>1.2B</td>
318
+ <td>61.46</td>
319
+ <td>75.5</td>
320
+ <td>61.8</td>
321
+ <td>47.07</td>
322
+ <td>49.31</td>
323
+ <td>81.8</td>
324
+ </tr>
325
+ <tr>
326
+ <td rowspan="4"><strong>General</strong><br><strong>VLMs</strong></td>
327
+ <td>Gemini-2.5 Pro</td>
328
+ <td>-</td>
329
+ <td>74.53</td>
330
+ <td>83.1</td>
331
+ <td><strong>77.2</strong></td>
332
+ <td>63.29</td>
333
+ <td>67.28</td>
334
+ <td>81.1</td>
335
+ </tr>
336
+ <tr>
337
+ <td>Qwen3-VL-235B</td>
338
+ <td>235B</td>
339
+ <td>70.91</td>
340
+ <td>83.3</td>
341
+ <td>74.8</td>
342
+ <td>54.63</td>
343
+ <td>59.43</td>
344
+ <td>82.1</td>
345
+ </tr>
346
+ <tr>
347
+ <td>Ovis2.6-30B-A3B</td>
348
+ <td>30B</td>
349
+ <td>59.34</td>
350
+ <td>60.2</td>
351
+ <td>65.8</td>
352
+ <td>52.03</td>
353
+ <td>57.00</td>
354
+ <td>64.4</td>
355
+ </tr>
356
+ <tr>
357
+ <td>GPT-4o</td>
358
+ <td>-</td>
359
+ <td>52.01</td>
360
+ <td>60.8</td>
361
+ <td>58.1</td>
362
+ <td>37.15</td>
363
+ <td>43.83</td>
364
+ <td>70.0</td>
365
+ </tr>
366
+ <tr>
367
+ <td rowspan="2"><strong>Pipeline</strong><br><strong>Tools</strong></td>
368
+ <td>Mineru2-pipeline</td>
369
+ <td>-</td>
370
+ <td>66.54</td>
371
+ <td>80.1</td>
372
+ <td>63.2</td>
373
+ <td>56.32</td>
374
+ <td>62.05</td>
375
+ <td>81.3</td>
376
+ </tr>
377
+ <tr>
378
+ <td>PP-StructureV3</td>
379
+ <td>-</td>
380
+ <td>60.32</td>
381
+ <td>78.2</td>
382
+ <td>53.7</td>
383
+ <td>49.07</td>
384
+ <td>62.06</td>
385
+ <td>79.1</td>
386
+ </tr>
387
+ </tbody>
388
+ </table>
389
+
390
+
391
+ ---
392
+
393
+ ## Evaluation Setup
394
+
395
+ 为保证评测公平性,请遵循以下设置:
396
+
397
+ 1. **Ground Truth**
398
+ - 使用 `ocr_data/md.zip` 中的 Markdown 文件。
399
+
400
+ 2. **Metric**
401
+ - 使用 `md2md` 评测指标。
402
+
403
+ > 注意:
404
+ > 虽然提供了 `md_original`,
405
+ > 但排行榜与标准评测请统一使用 `md/` 目录下的数据。
406
+
407
+ ---
408
+
409
+ # 📝 使用方式
410
+
411
+ ## 1. 下载并解压数据
412
+
413
+ ```bash
414
+ cd ocr_data
415
+ unzip pdf.zip
416
+ unzip md.zip
417
+ ```
418
+
419
+ ---
420
+
421
+ ## 2. 运行 OCR 模型
422
+
423
+ 对 `pdf/` 中的文档进行解析,并输出 Markdown 结果。
424
+
425
+ ---
426
+
427
+ ## 3. 执行评测
428
+
429
+ 将模型输出与 `md/` Ground Truth 进行比较。
430
+
431
+ ---
432
+
433
+ # 🔎 RAG Impact Evaluation
434
+
435
+ RAG Evaluation 用于评估 OCR 对端到端:
436
+
437
+ - Retrieval
438
+ - Generation
439
+ - QA
440
+
441
+ 效果的影响。
442
+
443
+ 相比字符级 OCR 指标,它更关注:
444
+
445
+ - 结构保留
446
+ - 语义保留
447
+ - 下游可用性
448
+
449
+ ---
450
+
451
+ ## RAG Evaluation Data
452
+
453
+ `RAG_eval/` 目录包含:
454
+
455
+ - `QA_pairs.jsonl`
456
+ - 共 2,071 条 QA 数据
457
+
458
+ - `doc_md/`
459
+ - 用于 RAG indexing 的 Ground Truth Markdown
460
+
461
+ ---
462
+
463
+ ## QA 数据格式
464
+
465
+ ```json
466
+ {
467
+ "doc_type": "cross_page_table",
468
+ "filename": "cross_page_table_1.md",
469
+ "title": "Document title",
470
+ "file_path": "RAG_eval/doc_md/cross_page_table_1.md",
471
+ "question_category": "...",
472
+ "question": "...",
473
+ "answer": "...",
474
+ "evidence": "..."
475
+ }
476
+ ```
477
+
478
+ ---
479
+
480
+ # ⚙️ RAG Pipeline Setup
481
+
482
+ 我们采用:
483
+
484
+ [FlashRAG](https://github.com/RUC-NLPIR/FlashRAG)
485
+
486
+ 中的 Naive Pipeline。
487
+
488
+ 配置如下:
489
+
490
+ | 模块 | 配置 |
491
+ |---|---|
492
+ | Embedding | BGE-M3 |
493
+ | Retrieval | Dense, Flat index, top-100 |
494
+ | Reranking | BGE-Rerank-V2-M3, top-10 |
495
+ | Generation | ChatGPT-5 |
496
+ | Chunking | HTML Tree Structure |
497
+ | Evaluation | RAGAS(GPT-OSS-120B Judge) |
498
+
499
+ ---
500
+
501
+ # 📈 RAG Evaluation Metrics
502
+
503
+ ### Context Recall
504
+
505
+ 评估检索结果是否包含 Ground Truth 证据。
506
+
507
+ ---
508
+
509
+ ### Answer Accuracy
510
+
511
+ 评估最终生成答案的正确性。
512
+
513
+ ---
514
+
515
+ # 🔥 关键 RAG 发现
516
+
517
+ | 文档类型 | OCR Accuracy | RAG Accuracy | Gap |
518
+ |---|---|---|---|
519
+ | VisualStyle | 82.9% | 52.8% | -30.1 pts |
520
+ | CrosspageTbl | 40.7% | 63.8% | +23.1 pts |
521
+ | UltraWide | 28.1% | 49.1% | low-low |
522
+ | MultiFont | 97.2% | 97.5% | ≈0 |
523
+
524
+ ---
525
+
526
+ **高 OCR Accuracy 并不意味着高 RAG 效果。**
527
+
528
+ 例如:
529
+
530
+ `VisualStyle` 文档虽然 OCR Accuracy 高达 82.9%,
531
+
532
+ 但由于 OCR 丢失:
533
+
534
+ - 删除线
535
+ - 颜色强调
536
+ - 视觉样式语义
537
+
538
+ 最终 RAG Accuracy 仅为 52.8%。
539
+
540
+ ---
541
+
542
+ # 📄 License
543
+
544
+ 本项目遵循开源协议发布。
545
+
546
+ 请在合法合规范围内使用本数据集。
547
+
548
+ 本数据集仅用于:
549
+
550
+ - 学术研究
551
+ - 非商业研究用途
552
+
553
+ ---
554
+
555
+ # 🙏 Acknowledgement
556
+
557
+ - 感谢 [OmniDocBench](https://github.com/opendatalab/OmniDocBench) 提供 OCR 评测代码。
558
+ - 感谢 [FlashRAG](https://github.com/RUC-NLPIR/FlashRAG) 提供 RAG Pipeline 框架。
559
+ - 感谢 [ragas](https://github.com/vibrantlabsai/ragas) 提供 RAG 评估代码。
560
+
561
+ ---
562
+
563
+ # 🖊️ Citation
564
+
565
+ 如果你使用了 InduOCRBench,请考虑引用:
566
+
567
+ ```bibtex
568
+ @misc{induocrbench,
569
+ title={When Good OCR Is Not Enough: Benchmarking OCR Robustness for Retrieval-Augmented Generation},
570
+ author={Lin Sun and Wangdexian and Jingang Huang and Linglin Zhang and Change Jia and Zhengwei Cheng and Xiangzheng Zhang},
571
+ year={2026},
572
+ eprint={2605.00911},
573
+ archivePrefix={arXiv},
574
+ primaryClass={cs.CV},
575
+ url={https://github.com/Qihoo360/InduOCRBench},
576
+ }
577
+ ```
assets/InduOCRBench_overview.png ADDED

Git LFS Details

  • SHA256: 72f2b217293c637dd43bdeb364c17c055b1ac7fcded572c3d0ae88935fe7558a
  • Pointer size: 132 Bytes
  • Size of remote file: 1.05 MB