Data Extraction of Circular-Shaped and Grid-like Chart Images

Area of Science:

• Computer Vision
• Image Processing
• Data Science

Background:

• Chart data extraction is vital for information recovery from visual data.
• Existing methods often lack standardized datasets, hindering result comparison.
• Publicly unavailable datasets limit research reproducibility and advancement.

Purpose of the Study:

• Develop a chart data extraction algorithm for circular and grid-like charts.
• Create a large-scale, publicly available dataset to facilitate research.
• Enable uniform comparison of results in chart data extraction.

Main Methods:

• Developed a novel, fully automatic low-level algorithm for chart data extraction.
• Utilized binary image processing instead of traditional pixel counting techniques.
• Created a large dataset of 120,000 chart images across 20 categories with ground truth.

Main Results:

• The proposed algorithm demonstrates effectiveness across diverse chart types.
• Achieved superior performance on a synthetic dataset, indicating state-of-the-art accuracy.
• Successfully extracted data from novel chart types like sunburst diagrams, heatmaps, and waffle charts.

Conclusions:

• A unified low-level approach is feasible for various chart types.
• The developed algorithm offers high accuracy and advances chart data extraction.
• The new dataset and algorithm will accelerate research and enable standardized comparisons.