Raster Plot Digitization

Scanned datasheets and bitmap-only PDF plots cannot be handled reliably by PDF vector extraction. The v1 workflow therefore uses a calibrated raster pipeline:

1. Prefer text, table, and vector extraction when available.
2. Render the target PDF page or plot region to a high-resolution bitmap.
3. Calibrate the plot with page number, plot rectangle, axis ranges, and log/linear axis modes.
4. Segment dark curve pixels with a tunable grayscale threshold.
5. Suppress long horizontal and vertical pixel runs, which are usually axes or grid lines.
6. Trace each requested X position by choosing the nearest continuous Y run.
7. Return points, pixel coordinates, coverage, gap count, and confidence for human review.

This is intentionally semi-automatic. It is more reliable to ask for a small amount of calibration than to silently guess axes from a scanned image. The GUI lets reviewers fill the page and plot rectangle from rendered evidence images, then run the raster digitizer for one curve at a time.

API

The local workbench exposes:

POST /api/raster-digitize

Example request:

{
  "session": "<upload-session>",
  "filename": "device.pdf",
  "page": 11,
  "plot_rect": [120, 180, 420, 360],
  "curve_name": "coss_pf",
  "x_range": [0.1, 1000],
  "y_range": [1, 100000],
  "threshold": 110,
  "initial_y_fraction": null
}

The response includes:

• points: calibrated X/Y values plus original pixel coordinates.
• confidence: a coverage/continuity score for review triage.
• metrics: requested point count, extracted point count, gaps, threshold, and dark-pixel fraction.

Reliability Notes

• Use the tightest possible plot rectangle. A large crop that includes legends, labels, or neighboring charts reduces confidence.
• For multi-curve plots, digitize one curve at a time. A future UI should add click-based seed points or color selection for Ciss, Coss, and Crss.
• If the scan has colored curves, use a curve-color segmenter before dark-pixel tracing. The current v1 path is optimized for black datasheet traces.
• OCR should be used as an assistant for axis tick suggestions, not as the sole source of calibration. Manual confirmation remains the reliable path.