Hiro-Layout: Document Layout Analysis for Patent and Technical PDFs

English | 简体中文

Hiro-Layout is a document layout analysis model for patent and technical PDF pages. It detects and classifies page regions such as text, titles, headers, footers, tables, formulas, chemical structures, figures, captions, search reports, bibliographies, and other patent-specific layout elements.

Highlights

• Patent-focused layout understanding: covers common patent PDF regions and patent-specific structures.
• Technical document coverage: evaluated on both patent PDFs and NPD PDFs.
• Fine-grained taxonomy: 25 layout categories across figure, text, and complex document elements.

Model Overview

Item	Details
Model name	Hiro-Layout
Current artifact	layout_model/RT-DETR_25.onnx
Task	Document layout analysis / page region detection
Input	Rendered PDF page image
Output	Layout regions with class labels
Domains	Patent PDFs, technical/NPD PDFs
License	Apache-2.0

Layout Taxonomy

Group	Class	Abbr.	Chinese
figure	graph	graph	图表
figure	drawing	draw	绘制图
figure	structure diagram	struc	结构图
figure	photograph	photo	照片
figure	table	tab	表格
figure	math equation	eqn	数学公式
figure	chemical formula	chem	化学式
figure	noise	noise	噪声
text	text	text	文本
text	title	title	标题
text	section title	sec	章节标题
text	page header	head	页眉
text	page footer	foot	页脚
text	marginal note	mnote	边注
text	caption	cap	说明
text	figure number	figno	编号
text	line number	lineno	行号
text	column number	colno	栏号
text	sequence	seq	序列表
complex	figure complex	figcx	图片组
complex	chemical reaction	rxn	反应式
complex	bibliography	bib	著录页
complex	search report	srep	搜索报告
complex	Table of Contents	toc	目录
complex	reference	ref	参考文献

Benchmarks

Metrics are reported as Precision, Recall, and F1.

Benchmark	Labels	Precision	Recall	F1
Patent PDF	33,054	0.8144	0.7711	0.7922
NPD PDF	17,769	0.7090	0.6983	0.7036

Patent PDF

#	Group	Abbr.	Class	Chinese	Labels	Precision	Recall	F1
1	figure	graph	graph	图表	215	0.7611	0.8000	0.7800
2	figure	draw	drawing	绘制图	420	0.8649	0.3048	0.4507
3	figure	struc	structure diagram	结构图	626	0.6579	0.8355	0.7361
4	figure	photo	photograph	照片	147	0.8378	0.8435	0.8407
5	figure	tab	table	表格	198	0.7759	0.9091	0.8372
6	figure	eqn	math equation	数学公式	399	0.7762	0.6692	0.7187
7	figure	chem	chemical formula	化学式	1,099	0.8792	0.8944	0.8868
8	figure	noise	noise	噪声	1,241	0.7025	0.7687	0.7341
9	text	text	text	文本	17,668	0.8182	0.8062	0.8122
10	text	title	title	标题	601	0.9117	0.8070	0.8561
11	text	sec	section title	章节标题	1,394	0.7968	0.7088	0.7502
12	text	head	page header	页眉	3,074	0.8187	0.7788	0.7983
13	text	foot	page footer	页脚	1,012	0.7432	0.6433	0.6896
14	text	mnote	marginal note	边注	421	0.7794	0.5202	0.6239
15	text	cap	caption	说明	80	0.6842	0.4875	0.5693
16	text	figno	figure number	编号	1,389	0.8955	0.7466	0.8143
17	text	lineno	line number	行号	341	0.7759	0.6598	0.7132
18	text	colno	column number	栏号	449	0.6964	0.4699	0.5612
19	text	seq	sequence	序列表	136	0.4430	0.2574	0.3256
20	complex	figcx	figure complex	图片组	1,416	0.8657	0.7373	0.7963
21	complex	rxn	chemical reaction	反应式	150	0.8898	0.7000	0.7836
22	complex	bib	bibliography	著录页	470	0.9615	0.7979	0.8721
23	complex	srep	search report	搜索报告	106	0.9052	0.9906	0.9459
24	complex	toc	Table of Contents	目录	0	0.0000	0.0000	0.0000
25	complex	ref	reference	参考文献	2	0.0000	0.0000	0.0000
ALL					33,054	0.8144	0.7711	0.7922

NPD PDF

#	Group	Abbr.	Class	Chinese	Labels	Precision	Recall	F1
1	figure	graph	graph	图表	248	0.6838	0.6976	0.6906
2	figure	draw	drawing	绘制图	9	0.0000	0.0000	0.0000
3	figure	struc	structure diagram	结构图	341	0.7454	0.7126	0.7286
4	figure	photo	photograph	照片	82	0.6071	0.6220	0.6145
5	figure	tab	table	表格	209	0.7533	0.8182	0.7844
6	figure	eqn	math equation	数学公式	298	0.6789	0.5604	0.6140
7	figure	chem	chemical formula	化学式	388	0.7324	0.8325	0.7793
8	figure	noise	noise	噪声	695	0.4823	0.4302	0.4548
9	text	text	text	文本	9,119	0.6943	0.7625	0.7268
10	text	title	title	标题	304	0.7130	0.5395	0.6142
11	text	sec	section title	章节标题	1,539	0.7337	0.6160	0.6697
12	text	head	page header	页眉	1,246	0.7464	0.7111	0.7283
13	text	foot	page footer	页脚	1,339	0.7711	0.6468	0.7035
14	text	mnote	marginal note	边注	190	0.5714	0.2947	0.3889
15	text	cap	caption	说明	573	0.8711	0.5899	0.7034
16	text	figno	figure number	编号	149	0.6078	0.4161	0.4940
17	text	lineno	line number	行号	41	0.6667	0.9268	0.7755
18	text	colno	column number	栏号	0	0.0000	0.0000	0.0000
19	text	seq	sequence	序列表	18	0.7000	0.3889	0.5000
20	complex	figcx	figure complex	图片组	734	0.7657	0.7480	0.7567
21	complex	rxn	chemical reaction	反应式	36	0.8947	0.4722	0.6182
22	complex	bib	bibliography	著录页	0	0.0000	0.0000	0.0000
23	complex	srep	search report	搜索报告	3	0.4286	1.0000	0.6000
24	complex	toc	Table of Contents	目录	76	0.8475	0.6579	0.7407
25	complex	ref	reference	参考文献	132	0.8148	0.3333	0.4731
ALL					17,769	0.7090	0.6983	0.7036

Usage

The current model artifact is an ONNX export:

layout_model/RT-DETR_25.onnx

Download the repository from the Hugging Face Hub and load the ONNX model with ONNXRuntime:

from pathlib import Path

from huggingface_hub import snapshot_download
import onnxruntime as ort

repo_dir = snapshot_download("PatSnap/Hiro-Layout")
model_path = Path(repo_dir) / "layout_model" / "RT-DETR_25.onnx"

session = ort.InferenceSession(str(model_path))
print("inputs:", [i.name for i in session.get_inputs()])
print("outputs:", [o.name for o in session.get_outputs()])

Use labels.json for the 25-class label mapping.

Repository Files

File	Purpose
README.md	Hugging Face model card in English
README_zh.md	Chinese model card
config.json	Model metadata used by Hugging Face Hub tooling and download statistics
EVALUATION.md	Detailed benchmark results derived from the workbook
labels.json	Machine-readable 25-class label mapping
layout_model/RT-DETR_25.onnx	ONNX model artifact
requirements.txt	Minimal dependencies for ONNX loading and image preprocessing
LICENSE	Apache-2.0 license
DISCLAIMER.md	Model limitations and responsible-use notes
NOTICE	Copyright and trademark notice
OPEN_SOURCE_CHECKLIST.md	Release checklist before public upload

Limitations

• Layout predictions may be inaccurate on low-resolution scans, heavily rotated pages, handwritten documents, unusual patent formats, or unseen page templates.
• Small objects and sparse categories can have unstable metrics when the evaluation set has very few labels.
• The model should not be used as the sole source of truth for legal, compliance, filing, archival, or customer-facing workflows without human review.
• Users are responsible for ensuring they have the right to process and share any documents used with this model.

License

This project is released under the Apache License 2.0. See LICENSE.

Copyright Notice

Copyright (c) 2026 Patsnap. All rights reserved except as expressly licensed under the applicable license terms.

Hiro-Layout, Hiro, Patsnap, and any associated names, logos, product names, service names, designs, and slogans are trademarks or registered trademarks of Patsnap or its affiliates. No trademark license is granted under the open source license or any model license unless expressly stated.