Analysis complete

Graph Neural Networks for Document Understanding

Document layout analysis remains a foundational task in computer vision and natural-language processing. Traditional CNN-based pipelines treat pages as flat pixel grids and miss the relational structure between visual regions.

2. Methodology

We model each page as a graph where nodes represent visual regions detected by a backbone CNN, and edges encode spatial proximity and reading-order relationships. A two-layer Graph Attention Network propagates context.

Method	F1	mAP
CNN baseline	0.891	0.842
DOCUGRAPH (ours)	0.974	0.928

[Figure 1 — Pipeline overview]

Results on the PubLayNet benchmark show consistent improvement across all region categories, with the most significant gains on tables and figures.

{
  "document_id": "doc_20260507_a8f2",
  "pages": 1,
  "confidence": 0.968,
  "regions": [
    {
      "type": "header",
      "bbox": [40, 38, 660, 84],
      "text": "Graph Neural Networks for…",
      "confidence": 0.992,
      "level": 1
    },
    {
      "type": "paragraph",
      "bbox": [40, 100, 660, 178],
      "text": "Document layout analysis…",
      "confidence": 0.978
    },
    {
      "type": "header",
      "bbox": [40, 200, 200, 234],
      "text": "2. Methodology",
      "confidence": 0.985,
      "level": 2
    },
    {
      "type": "table",
      "bbox": [40, 320, 660, 410],
      "rows": 3,
      "cols": 3,
      "confidence": 0.951
    },
    {
      "type": "figure",
      "bbox": [40, 430, 660, 540],
      "caption": "Figure 1 · Pipeline overview",
      "confidence": 0.947
    }
  ]
}

# Graph Neural Networks for Document Understanding

Document layout analysis remains a foundational task
in computer vision and natural-language processing.
Traditional CNN-based pipelines treat pages as flat
pixel grids and miss the relational structure between
visual regions.

## 2. Methodology

We model each page as a graph where nodes represent
visual regions detected by a backbone CNN, and edges
encode spatial proximity and reading-order relationships.

| Method            | F1    | mAP   |
|-------------------|-------|-------|
| CNN baseline      | 0.891 | 0.842 |
| DOCUGRAPH (ours)  | 0.974 | 0.928 |

![Figure 1 — Pipeline overview](fig1.png)

Results on the PubLayNet benchmark show consistent
improvement across all region categories.