Tesseract.js · SVG regions · in-browser

Word Finder

Scan an image for words, then fix what the scanner got wrong — box or trace anything it missed, delete anything it invented. Every region is an SVG path you can copy.

Drop an image here

or click to browse · paste works too

V select · R box · P pen · T trace · E remove · ⏎ close path · ⌫ delete selected · Ctrl/⌘ Z undo · Esc cancel — the image never leaves your browser. powered by 3gem research and insights

How this works

everything runs locally in your browser

The basics

Load an image, press Find words, and Tesseract.js reads it — the file is never uploaded anywhere. The first scan downloads a language model of a few megabytes, so it is slower than the ones after it. Each word it finds becomes a numbered, clickable region on the image and a row in the list. Click any word in the list to rename it if the scanner guessed wrong.

Tools

Select · V
Click a region to select it and read its coordinates. ⌫ removes it.
Box · R
Drag a rectangle around a word the scanner missed.
Pen · P
Click to drop points one at a time. With three or more down, click the first point again — it turns yellow when you are close enough to snap — or press ⏎ to close the shape.
Trace · T
Draw freehand around a word; the stroke is smoothed and closed for you.
Remove · E
Click regions to delete them. The ✕ in each list row does the same.

Anything you add by hand is read on its own: the region is cropped out, enlarged, and run through OCR in single-line mode, which is far more accurate on one word than a full-page pass. Hand-drawn regions are blue, scanner results yellow.

Coordinates

Every region — found or drawn — is an SVG path in the image's own pixel space, so the numbers work against the original file at any display size.

  • Box and scanner hits: M291,233 H574 V292 H291 Z
  • Pen: M291,233 L574,231 L576,290 Z
  • Trace: relative smoothed curves, m291,233c-1,0 -3,59 …z

Copy d takes the selected path. Copy all takes every path currently passing the filters, joined with |. The SVG export writes a ready-made overlay where each path carries an id, a data-word attribute and a tooltip.

Filters

Min confidence hides scanner results it was less sure about — useful for clearing noise off a busy scan. Hand-drawn regions have no score, so the slider never hides them. The search box filters by text. Both affect what is drawn, listed and copied by Copy all; CSV, JSON, SVG and Copy list always take everything.

Shortcuts

V select · R box · P pen · T trace · E remove
⏎ close pen path · ⌫ delete selected · Esc cancel or deselect
Ctrl/⌘ Z undo · Ctrl/⌘ ⇧ Z redo · ? this window

If a scan goes badly

  • Sharp, high-contrast, upright text reads best; photos of screens and skewed pages read worst.
  • Small text often fails at low resolution — a larger version of the same image usually fixes it.
  • Set the language before scanning; the wrong model produces confident nonsense.
  • If a word comes back as untitled, the crop was unreadable. Type the word in yourself; the coordinates are still good.