Skip to content
DemoFAQ

PDF Generation Template Editor

Open-source visual PDF generation engine with customizable templates and developer-friendly APIs.

View on GitHub

Table Recognition ​

Overview ​

Table Recognition reconstructs the internal structure of tables detected during layout analysis, including rows, columns, merged cells, and cell boundaries, so that the converted document preserves the original tabular semantics instead of producing a flat grid of text fragments.

It is controlled by the independent option enable_ai_table_recognition, which is enabled by default. The table model is only invoked for table regions reported by layout analysis whose detection confidence is below the trusted threshold; high-confidence native PDF tables bypass the model to save inference time.

Typical scenarios that benefit from Table Recognition:

  • Borderless or partially bordered tables, where ruling lines alone cannot describe the structure.
  • Tables with merged header cells, multi-row headers, or spanning cells, such as financial statements, lab reports, and invoices.
  • Scanned tables processed by OCR, where geometric reconstruction is required before cell-level data extraction.

Features that support Table Recognition:

  • PDF to Word
  • PDF to Excel
  • PDF to PowerPoint (Slides)
  • PDF to HTML
  • PDF to RTF
  • PDF to CSV
  • Extract PDF to JSON
  • Extract PDF to Markdown

Notice ​

  • Table Recognition runs only when layout analysis is active (i.e. enable_ai_layout = true, or implicitly when enable_ocr = true).
  • You need to load the DocumentAI model before using Table Recognition, or plug in your own table model via the callbacks described in 3.12 Use Custom AI Models via Callbacks.
  • Setting enable_ai_table_recognition = false disables the table model entirely; detected table regions will then fall back to geometric reconstruction from the underlying page objects.
  • The number of Table Recognition model instances can be tuned via the second parameter of LibraryManager::SetDocumentAIModelCount. See 3.2 Set DocumentAI Model.
  • The model profile used by Table Recognition is selected with table_recognition_accuracy; see "Table Model Profile" below.

Table Model Profile ​

Table Recognition provides two model profiles, selected per conversion task through ConvertOptions.table_recognition_accuracy (type TableRecognitionAccuracy):

ProfileDescription
e_HighAccuracy (0)Prioritizes recognition accuracy and may use more resources. The default behavior in 4.2.0
e_Balanced (1, default since 4.3.0)Balances recognition accuracy, resource usage, and throughput
  • Default changed: 4.2.0 used the high-accuracy profile by default, while 4.3.0 uses the balanced profile by default. After upgrading without changing any code, table recognition results and resource usage will differ from 4.2.0.
  • The balanced profile only needs the base package documentai.model. The high-accuracy profile requires the table extension package table_legacy.model to be loaded as well (see 3.2 Set DocumentAI Model).
  • Models of both profiles can be loaded at the same time; each conversion task uses the profile given by its own table_recognition_accuracy.
  • If the requested profile has no corresponding model (for example, the high-accuracy profile without table_legacy.model), the conversion still succeeds, but tables are recognized with the rule-based method instead, so the result differs from that profile.
  • A profile name does not guarantee better results for every document.
c++
LibraryManager::SetDocumentAIModel("path/documentai.model", -1);
LibraryManager::SetDocumentAIModel("path/table_legacy.model", -1);

ConvertOptions opt;
// Use the high-accuracy profile (requires table_legacy.model to be loaded).
opt.table_recognition_accuracy = TableRecognitionAccuracy::e_HighAccuracy;
CPDFConversion::StartPDFToWord("word.pdf", "password", "path/output.docx", opt);

Sample ​

This sample demonstrates how to convert a PDF to a DOCX file with Table Recognition enabled.

c++
// Set the DocumentAI model path.
LibraryManager::SetDocumentAIModel("path/documentai.model");

ConvertOptions opt;
// Layout analysis must be enabled for Table Recognition to take effect.
opt.enable_ai_layout = true;
// Enable AI table recognition (enabled by default; set to false to disable).
opt.enable_ai_table_recognition = true;
CPDFConversion::StartPDFToWord("word.pdf", "password", "path/output.docx", opt);