PDF OCR Layer Alignment Checker: Does the Text Sit on the Words?

What it does. For each page the tool renders the page (hidden text is not drawn, so the render shows the scanned picture), reads the boxes of the text layer from the PDF text matrices, and measures how much dark pixel content lies under each box. It then tries moving the whole layer around the page and turning it by 90, 180 and 270 degrees to see whether another position fits clearly better.

The scoring, defined. Nothing is invented; the numbers have these meanings:

• Ink: a pixel at or below the automatic black-and-white threshold (Otsu's method) of the rendered page.
• Text-on-ink rate: the share of text boxes that hold at least half of the typical ink density the boxes have at their best-fitting position (never less than 4% or more than 15% ink). This is the page's alignment score. It is not an OCR accuracy.
• Offset: reported only when a shift of at least 3 pixels raises the fit clearly (fit of the stored position is 85% or less of the best), and given in millimetres and direction.
• Rotated: reported only when turning the layer raises the share of boxes on ink by at least 25 points and to at least 50%.
• Inconclusive: when boxes land on ink about equally well everywhere (dense or even pages), the tool says it cannot tell instead of guessing.

What you see. A page table with status (aligned, offset, rotated, partly aligned, poor fit, inconclusive, not applicable), the score and the reason in words, and an overlay view of any page: green boxes sit on ink, red boxes do not, and for an offset page the better position can be drawn in blue. Pages that are not scanned pictures with hidden text (digital text, no picture, a blank render) are marked not applicable with the reason.

Standards. The hidden text of OCR output uses text rendering mode 3 (ISO 32000-1, 9.3.6), which is how the tool tells hidden text from visible text.

Read more. The guide The searchable PDF and its OCR layer explains how the hidden text is built and how it fails.

PDF files up to 80 MB, the first 1, 30 or 100 pages as chosen. Works on pages that carry a full-page picture and a hidden text layer. Password-protected PDFs are refused with a message.

1. Drop a searchable (OCR'd) PDF onto the box and choose how many pages to check.
2. Read the verdict and the page table. Pages needing attention are listed with the reason.
3. Open the overlay of a page and look at where the boxes fall; for an offset page tick the option to see the better-fitting position.
4. Re-run OCR with the right page orientation or settings, or use the program that made the layer, and check the new file.

• Search highlights in a scanned contract appear next to the words
• Selecting text in a searchable PDF selects the wrong lines
• Checking the output of an OCR program before archiving a scan
• Finding pages whose OCR layer is rotated or shifted in a long scan
• Telling an OCR layer that is misplaced from one that is merely inaccurate

The method compares positions of boxes with dark pixels. It assumes dark text on a light background in mostly horizontal lines; handwriting, photographs, very faint or heavily skewed scans and negative images reduce its reliability, and the tool then reports inconclusive or not applicable. Skew of a few degrees can lower the score without being reported as an offset or rotation. Pages are rendered at a moderate resolution, so offsets smaller than about a millimetre are not resolved.

It measures overlap, not recognition: it cannot say whether the recognised words are spelled correctly, and it does not change the PDF. Rendering uses the pdf.js reader in your browser.