Document layout analysis is the detection of a page's structural regions, its tables, headers, paragraphs, and figures, before any text is read for meaning. In commercial real estate, it locates the rent grid on a rent roll and the clause blocks in a lease, so downstream steps know where each value lives. It is the map that tells extraction where to look.
How Does Document Layout Analysis Work?
Document layout analysis works by treating a page as an object detection problem: a model draws bounding boxes around regions and labels each by type. It is often the first task in a document understanding system, breaking a page into identifiable sections, per DocLayNet research. Deep learning models trained on large annotated datasets have proven effective at this segmentation.
The reference benchmark, DocLayNet from IBM Research, contains 80,863 manually annotated pages and labels 11 region classes: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, and Title. Performance is measured in mean Average Precision (mAP) using the COCO metric. Layout analysis runs before or alongside optical character recognition, so the system reads each region with logic suited to its type.
Region type | What it holds in a CRE document |
Table | The rent grid on a rent roll or the pro forma on an OM |
Section-header | Clause titles in a lease, such as "Rent" or "Options" |
Text | The paragraph body of a lease clause |
Page-header/footer | Running titles, page numbers, and boilerplate to ignore |
Picture | Site plans, maps, and photos in an offering memorandum |
Why Document Layout Analysis Matters
Document layout analysis matters because a value's meaning depends on where it sits, and reading a page as one flat stream loses that structure. A number inside a rent-roll table is a monthly rent; the same number in a footer is a page count. Layout analysis is what separates the table from the boilerplate before extraction assigns meaning.
Accuracy varies sharply by document diversity, which is the operator's real problem. Recent methods reach 97.3% mAP on PubLayNet, whose pages come only from scientific articles, but 81.6% mAP on the far more varied DocLayNet, per published benchmarks. Commercial real estate documents resemble the harder case: leases, rent rolls, and offering memorandums have inconsistent, custom layouts, so a value in a nonstandard table is where layout errors and downstream extraction errors begin.
Example
Document layout analysis is easiest to see when a page has mixed regions. A rent-roll page contains 10 distinct regions: one large table, two headers, six text blocks, and a footer. The model draws a bounding box for each and labels its type, and the pipeline checks that the table was found before extraction runs.
Metric | Value |
Regions on the page | 10 |
Regions detected | 10 |
Regions labeled correctly | 9 |
Region label accuracy | 90% |
Table region detected | Yes |
Mislabeled regions flagged | 1 |
The model detects all 10 regions and labels 9 correctly, a 90% label accuracy, with one text block misread as a footer. Critically, the table region is found, so the rent grid routes to table extraction intact. The single mislabeled block is flagged for review. If layout analysis had missed the table boundary, every rent value inside it would extract into the wrong structure, which is why table detection is the gate the rest of the pipeline depends on. The region count and accuracy rate here are illustrative inputs.
Variations and Edge Cases
Document layout analysis behaves differently by page complexity and how standard the layout is. A single-column lease page is straightforward; a multi-column OM with overlapping graphics is not. The variants below shift detection accuracy.
Variant | Treatment |
Single-column text | Simple region order; high detection accuracy |
Multi-column layout | Reading order must be reconstructed across columns |
Dense tables | Nested rent grids; table boundaries can blur |
Mixed media | Text wrapped around maps or photos in an OM |
Scanned and skewed | Degraded input lowers boundary precision |
Document Layout Analysis vs Optical Character Recognition
Document layout analysis is often confused with optical character recognition, but they answer different questions about a page. Optical character recognition reads what the characters are, converting pixels into text. Document layout analysis reads where the regions are, marking this block a table and that block a header, without reading a single word for meaning.
Layout analysis finds the structure; optical character recognition reads the content inside it. They are complementary steps: knowing a region is a table tells the pipeline to send it to table extraction, and knowing the characters inside tells it what the cells say. Structure first, then text.
Frequently Asked Questions
What is document layout analysis?Document layout analysis is the detection of a page's structural regions, such as tables, headers, text blocks, and figures, before any text is read for meaning. It draws labeled bounding boxes around each region so downstream extraction knows where each type of content sits on the page.
How accurate is document layout analysis?Recent methods reach 97.3% mAP on the PubLayNet benchmark of scientific articles but 81.6% mAP on the more varied DocLayNet dataset, per published benchmarks. Diverse, nonstandard layouts like CRE leases and rent rolls fall closer to the harder case.
What is the difference between document layout analysis and OCR?Document layout analysis detects where regions are on a page and labels them, such as table or header, without reading words. Optical character recognition reads the characters inside those regions. Layout analysis finds structure; OCR reads content, and both feed extraction.
Related Terms
Optical Character Recognition
Table Extraction
Intelligent Document Processing
Field Extraction
Structured Data Extraction