Vision Text

Recognise text in images with Apple Vision OCR.

What it does

Vision Text reads an image and finds any text in it, using Apple's on-device Vision framework. For each piece of text found it reports the words, where they sit in the image, and a confidence, plus a single string joining everything found and a count of how many regions turned up.

It runs live: fed a camera or screen capture, it keeps re-reading the image as it updates, with no separate cook step to wire.

When to use it

ParameterTypeDefault
imageOneOf([FieldOf(Raster)])
levelString"accurate"
confidence_thresholdNumber0.300
language_correctionBooleantrue
enabledBooleantrue

Gotchas

This node relies on Apple's Vision framework and works on macOS only. On other platforms it does not appear in the palette at all.

Language Correction is on by default, which lets Vision correct recognised words against dictionary guesses. That suits prose, but it can silently rewrite exact strings (codes, serial numbers, unusual names) into something that reads better and is not what was on screen. Turn it off for literal transcription.

Worked example

  1. Add a Vision Text node and wire an image, from a Camera, Screen Capture, or any image output, into Image.

  2. Leave Level on "accurate" for still frames, or switch to "fast" for a live camera feed needing lower latency.

  3. Read Full Text for a joined string, or wire Text Regions into a downstream node to work with individual words and their positions.

  4. Raise Confidence Threshold if junk regions are turning up from background clutter.

See also

Vision Analyze · Object Detect · Screen Capture · Collections