Chart Data Extraction

Chart data extraction recovers numerical values, series, and labels from plots so information locked in images can be reused and analyzed. Approaches include OCR and coordinate-based digitizers with manual or automatic axis calibration, specialized chart models, and vision-language models, sometimes aided by spatial prompts, synthetic training data, or self-ensembling. Benchmarks test chart families and settings ranging from ChartQA-style digital charts to dense, hand-drawn, and domain-specific figures, using measures of value error, series alignment, and usability. Models can often reconstruct table structure more reliably than precise values, and performance remains sensitive to chart type, visual density, and domain, making robust evaluation and human verification important.