Skip to content

How It Works

Jump to bottom
Nick edited this page Dec 18, 2025 · 1 revision

How It Works

Chart2CSV uses computer vision and OCR to extract data from chart images.

Pipeline Overview

Image → Preprocess → Detect Axes → OCR Labels → Extract Points → Output CSV

Step by Step

1. Image Preprocessing

  • Convert to grayscale
  • Enhance contrast
  • Remove noise

2. Plot Area Detection

  • Find the chart boundaries
  • Crop to the actual plot region
  • Ignore titles, legends, and margins

3. Axis Detection

  • Find X and Y axis lines using line detection
  • Determine axis positions in pixels

4. Tick Label OCR

Two backends available:

Tesseract (offline):

  • Free, runs locally
  • Good for clear, high-resolution images

Mistral Vision (cloud):

  • Uses AI to read text
  • Better accuracy, especially for difficult fonts
  • Requires API key

5. Coordinate Transformation

  • Map pixel positions to actual values
  • Uses detected tick labels to build the mapping
  • Handles linear and logarithmic scales

6. Data Point Extraction

Depends on chart type:

Scatter plots: Detect colored dots using blob detection

Line charts: Trace the line using skeletonization

Bar charts: Detect rectangles and measure heights

7. Output

  • CSV file with x,y coordinates
  • JSON metadata (confidence, warnings, parameters)
  • Visual overlay showing detected points

Confidence Scoring

Chart2CSV reports a confidence score (0.0 to 1.0):

Score Meaning Action
≥ 0.7 High confidence Trust the results
0.4 - 0.7 Medium Check the overlay
< 0.4 Low Use manual calibration

Caching

OCR results are cached to speed up repeated extractions:

  • Cache location: ~/.cache/chart2csv/ocr/
  • Separate caches for Tesseract and Mistral
  • Use --no-cache to bypass

Clone this wiki locally