taulu 3.0.1

Segment a table from an image

Segmentation of tables from images

Documentation · Changelog · Examples · Colab demo

[!WARNING] v3.0 renamed several classes and parameters. See the migration guide and run uv run -m taulu.migrate config.toml --inplace to upgrade TOML configs.

Features

• Rust core (PyO3) for grid growing, A* pathfinding, row detection
• Two-page tables via Split — different headers and parameters per side
• Auto row heights detected from cross-correlation peaks
• TOML config with JSON Schema for editor autocomplete
• Notebook UI for header annotation and cell inspection (matplotlib + ipywidgets)
• OpenCV debug view for live parameter tuning

Data Requirements

Images of tables with clearly visible rules (cell borders).

For the automated workflow, tables should include a recognizable header — used to locate the first cell and infer column widths.

Tables should be roughly axis-aligned. Minor warping is fine.

Installation

pip install taulu
# or
uv add taulu

Quickstart

from taulu import Taulu

Taulu.annotate("table.png", "header.png")          # one-time, interactive
taulu = Taulu("header.png")
grid = taulu.segment_table("table.png")
grid.show_cells("table.png")                        # click cells to inspect

For two-page (split) tables, see examples/example.py.

Workflow

Taulu orchestrates these components:

Class	Role
TemplateMatcher	Locate the header in the image (ORB / SIFT / AKAZE)
TableTemplate	Header annotation: column rules + expected cell sizes
TableDetector	Find rule intersections (binarization → morphology → cross-kernel)
SegmentedTable	Output grid: cell lookups, cropping, persistence

Annotation is two clicks per line:

Parameters

Most-tuned Taulu(...) parameters:

Parameter	Purpose
template_path	Header image (.png + .json annotation). Pass Split(...) for two-page tables.
intersection_kernel_size, line_thickness	Shape the cross-kernel to match real corner geometry after morphology.
line_gap_fill	Dilation size — bridges broken rules. Affects optimal line_thickness.
search_radius	Search window around predicted corner. Larger = more warp tolerance, more false positives.
binarization_sensitivity	Sauvola threshold. Higher = more aggressive noise removal.
row_height_factor	Row height as fraction of header height. Float or list.
auto_row_heights	Detect per-row heights from cross-correlation peaks.

Cross-kernel shape vs. search region:

   

[!TIP] Run with debug_view=True to see binarization, morphology, and search regions live — easiest way to tune.

Full parameter docs: Taulu reference.

SegmentedTable methods

Method	Purpose
save / from_saved	Persist grid to/from JSON
cell((x, y))	Pixel → (row, col)
cell_polygon((r, c))	Cell → 4-corner polygon
region(start, end)	Bounding polygon over a cell range
crop_cell / crop_region	Perspective-correct crop
highlight_all_cells	Render all cell outlines
show_cells	Interactive click-to-inspect (OpenCV / notebook)

Configuration via TOML

"$schema" = "./taulu-config.schema.json"
template_path = "header.png"
binarization_sensitivity = 0.05
intersection_kernel_size = 41

[search_radius]                   # per-side override (split tables)
left = 60
right = 80

from taulu import Taulu, TauluConfig
taulu = Taulu.from_config(TauluConfig.from_toml("config.toml"))

Generate the schema for editor autocomplete:

uv run -m taulu.schema > taulu-config.schema.json

Citation

@software{taulu,
  author  = {Peeters, Miel and {GhentCDH}},
  title   = {taulu: segmentation of tables from images},
  url     = {https://github.com/ghentcdh/taulu},
  version = {3.0.0},
  year    = {2026}
}

See CITATION.cff.

---

Development by Ghent Centre for Digital Humanities — Ghent University.
Funded by the GhentCDH research projects, within the CLARIAH-VL consortium.