trace-digitiser 0.5.0

Template-free computer-vision digitisation pipeline for raster scientific line plots.

trace-digitiser

Template-free computer-vision digitisation pipeline for raster scientific line plots.

Given a raster image of a scientific figure (PNG, JPEG, etc.), this tool detects plot panels, reads y-axis tick labels via OCR, segments coloured traces, and exports digitised data as CSV files — all without requiring manual calibration points, known trace colours, or panel coordinates. The user supplies only a high-level layout hint (e.g. "stacked", "horizontal", "grid").

Installation

Requires Python 3.10+ and Tesseract OCR.

# Install Tesseract (Ubuntu/Debian)
sudo apt-get install tesseract-ocr

# Install Tesseract (macOS)
brew install tesseract

# Install the package
pip install -e .

# With development dependencies
pip install -e ".[dev]"

Quick start

Python API

from trace_digitiser import digitise

result = digitise(
    "figure.jpg",
    layout_mode="stacked",
    expected_rows=2,
    expected_cols=1,
    output_dir="outputs",
)

# Digitised traces as a DataFrame
print(result.trace_data.head())

# Panel metadata
for panel in result.panels:
    print(f"Panel {panel.panel_id}: {panel.width}×{panel.height}px, "
          f"calibration={panel.calibration.scale_type}")

Command line

# Single image
trace-digitiser figure.jpg --layout stacked --rows 2 --cols 1 -o outputs/

# Batch processing
trace-digitiser figures/*.jpg --layout auto -o results/

# Generate synthetic test figures
trace-digitiser --generate-test-figures -o test_figures/

Layout modes

Mode	Use case	Example
single	One plot panel	--layout single
stacked	Vertically stacked panels	--layout stacked --rows 2 --cols 1
horizontal	Side-by-side panels	--layout horizontal --rows 1 --cols 3
grid	Row×column subplot grid	--layout grid --rows 2 --cols 2
auto	Unconstrained detection	--layout auto

Output files

For input figure.jpg with default settings:

File	Description
figure_automated_digitised_trace_by_column.csv	One row per x-pixel column per trace, with center_estimate, upper_envelope, lower_envelope
figure_automated_panel_metadata.csv	Panel coordinates, calibration parameters, detected x-labels
figure_automated_digitised_summary_by_label.csv	Summary statistics per detected x-label interval (written only if labels are found)

Project structure

trace_digitiser/
├── pyproject.toml
├── src/trace_digitiser/
│   ├── __init__.py          # Public API: digitise()
│   ├── models.py            # Dataclasses: Panel, Calibration, Trace, DigitiserResult
│   ├── io.py                # Image loading, CSV output
│   ├── geometry.py          # Box area/IoU/containment helpers
│   ├── line_detection.py    # Horizontal and vertical line detectors
│   ├── panel_detection.py   # Candidate generation, deduplication, layout selection
│   ├── calibration.py       # Y-axis OCR calibration (linear + log) and cross-row propagation
│   ├── x_calibration.py     # X-axis OCR calibration (numeric ticks)
│   ├── ocr.py               # Tesseract OCR wrappers for tick and x-label reading
│   ├── trace_detection.py   # HSV/CIELAB colour segmentation + achromatic trace detection
│   ├── digitise.py          # Column-by-column trace digitisation
│   ├── summarise.py         # Interval-label summarisation
│   ├── diagnostics.py       # Debug overlays and diagnostic file output
│   ├── synthetic.py         # Synthetic test-figure generation
│   └── cli.py               # Command-line interface
└── tests/
    ├── conftest.py           # Shared fixtures (synthetic images)
    ├── test_geometry.py
    ├── test_panel_detection.py
    ├── test_trace_detection.py
    └── test_integration.py   # Panel count, calibration quality, trace RMSE

Development

# Run tests
pytest

# Run tests with coverage
pytest --cov=trace_digitiser

# Lint
ruff check src/ tests/

Limitations

This tool digitises visible pixels from raster images. It does not recover original raw data. Key limitations:

• Linear and log y-axes supported — broken, dual, and other nonlinear axes are not supported yet.
• X-axis calibration is best-effort — numeric x-ticks are OCR'd when possible; otherwise x-values are normalised 0–1 or labelled by interval.
• Black/grey traces partially supported — achromatic trace detection is available but works best when traces have enough contrast against the background.
• Requires visible structure — axes, borders, or gridlines must be present for panel detection.
• OCR fragility — small, rotated, or low-contrast tick labels may fail.

Changelog

v0.1.0

Refactored from notebook to installable package with 15 modules, dataclasses, CLI, and test suite.

Panel detection improvements:

• Hint-aware split_y_clusters — uses the user's expected_rows to split evenly-spaced gridlines at the largest gaps, even when the inter-panel gap is only marginally wider than intra-panel gaps.
• Post-hoc panel subdivision — apply_layout_hint can split oversized candidates at their gridlines when initial detection yields fewer panels than expected.
• Improved grid detection — properly cross-matches row/column structure.

Trace detection improvements:

• CIELAB clustering mode (use_lab=True) — clusters non-background pixels in perceptually uniform colour space using mini-batch k-means.
• Achromatic trace detection — a separate pass detects black/grey traces by looking for low-saturation, horizontally continuous structures distinct from grid/axis lines.

Calibration improvements:

• Log-scale y-axis detection — when OCR'd tick values are better explained by log10(y_value) = a * y_pixel + b, the calibration uses log scale automatically.
• X-axis calibration — OCRs numeric x-axis ticks, fits x_value = a * x_pixel + b, and adds x_value column to the trace CSV.

Infrastructure:

• Diagnostic overlays save to files via save_diagnostics=True.
• Tesseract error handling — graceful recovery from OCR crashes on degenerate crops.
• No Colab dependency.
• Quantitative test suite with panel count, calibration quality, and trace RMSE metrics against synthetic ground truth.