Tiny Online Tools
The layer nobody can see
When a scan goes through OCR, the recogniser does not replace the picture. It leaves the scanned image exactly as it was and writes the words it recognised on top of it in PDF render mode 3 — "neither fill nor stroke". The text is there, selectable and searchable, but no viewer will ever paint a single pixel of it.
That is a sensible design, and it makes OCR quality impossible to judge. A search that returns nothing tells you the word is missing; it does not tell you whether the recogniser read Received as Recewed, or shifted an entire column one line up.
What this tool shows you
Every recognised word is drawn back into the exact box it was recognised from, over the scan it came from. Four views:
- Boxes only — outlines, so you can see coverage and alignment without the text getting in the way.
- Text over image — the recognised words drawn semi-transparently over the scan, with an opacity slider. Where recognition drifted, the drawn word slides off the scanned word underneath.
- Text only — the layer alone on blank paper, which is what a screen reader effectively gets.
- Image only — the scan as it was, for comparison.
Switch between word, line and block boxes, search the recognised text and see matches highlighted on the page, and click any box to jump to it in the list.
Judging quality
Per page you get the word and character counts, how much of the text is genuinely invisible, the average word box size, and a coverage figure — how much of the page area the OCR boxes actually cover. A recogniser that gave up halfway down the page shows as low coverage long before you notice the missing words.
Words whose shape is statistically odd are flagged: no vowels, letters mixed with digits, single-character fragments, symbol runs, boxes badly out of step with their neighbours. These are suspicions, not errors — a part number legitimately mixes letters and digits.
Export
Plain text, JSON with every bounding box and font, or hOCR-style HTML for downstream tooling. Coordinates are PDF points with the origin at the top-left.
Privacy
Nothing is uploaded. The PDF is parsed, rendered and measured entirely inside your browser.