OCR can read visible labels; it cannot recover exact unlabeled values or determine what a chart proves. The workflow keeps directly read values separate from visual estimates.
This route recognizes visible text and numbers while keeping estimates clearly separate from source values and refusing to infer what the chart proves.
Recognition is only the first stage. Use Numeric OCR Validation for Dates, Totals and Identifiers for the relevant control checks and Table OCR Quality Control: Rows, Columns and Merged Cells for values that must be exact. If the output structure changes, compare Table to Markdown with OCR Quality Checker.
Trace every label and value to visible evidence
| Review point | Likely failure | Validation action |
|---|---|---|
| Scale | A truncated, logarithmic, or reversed axis is missed | Record scale type, baseline, ticks, and units first |
| Series | Legend labels detach from colors or marks | Map each series visually before entering values |
| Estimates | Interpolated values are presented as exact | Label every value as read, calculated, or estimated |
Record the scale and series before entering values
Identify chart type, units, scale, and zero baseline.
Match each legend entry to its visual series.
Label values as read, calculated, or estimated in the reconstructed dataset.
Chart reconstruction scenario
For a bar chart with labels on only some bars, transcribe those printed labels as read values. If another bar must be estimated between ticks, record the estimate method and uncertainty in a separate field. Do not give estimated values the same status as visible numbers.
When underlying data or a digitization tool is needed
Use the source dataset when it is available. Use a chart digitization tool for coordinate-based estimates, then preserve the axis calibration and uncertainty rather than describing the result as exact OCR output.