Table Extraction

Approach

EdgeParse uses two complementary methods for table detection:

1. Border-Based Detection

When PDFs contain ruling lines (re, l, m operators), EdgeParse identifies grid patterns and maps text runs into cells.

2. Cluster-Based Detection

For borderless tables, EdgeParse analyzes text alignment patterns to infer column boundaries and row groupings.

Cell Merging

After initial detection, EdgeParse identifies spanning cells by analyzing:

• Cells with content that overlaps multiple grid positions
• Header rows with merged cells
• Row/column span attributes from tagged PDF structure

TEDS Score

EdgeParse achieves a TEDS score of 0.783 — the highest among rule-based tools:

Tool	TEDS Score	Type
Docling	0.887	ML-based
Marker	0.825	ML-based
EdgeParse	0.783	Rule-based
EdgeQuake	0.795	ML-enhanced
PyMuPDF4LLM	0.540	Rule-based

Output Format

Tables in JSON output include full cell structure:

{
  "type": "table",
  "id": 3,
  "page number": 1,
  "rows": [
    {
      "row number": 0,
      "cells": [
        {"col": 0, "content": "Quarter", "is header": true},
        {"col": 1, "content": "Revenue", "is header": true}
      ]
    }
  ]
}