OCR on scanned manuals: what goes wrong

Optical character recognition (OCR) turns a scanned page into text a machine can search, and on a maintenance manual its mistakes are not cosmetic. A misread character in a part number or a torque value produces something that looks valid and is wrong. OCR fails in a few recurring ways, most of them traceable to the scan itself, and the engine can report how sure it was of every word. That report is the signal worth keeping.

Where the errors start

An OCR engine does not read a page in one step. Tesseract, an open-source engine, documents its stages in a quality guide: it converts the image to black and white, finds the lines of text, and only then recognizes characters. A mistake made early is handed to the next stage as if it were correct.

The guide ties four scan conditions to those stages:

• Resolution. The engine works best on images of at least 300 dots per inch.
• Background and noise. An unevenly dark page degrades the conversion to black and white, and some noise cannot be removed at that step.
• Skew. A tilted scan breaks line finding. The page should be rotated until the text lines are horizontal.
• Borders. Dark edges left by a scanner can be read as characters.

These notes describe one engine. A study that compared several OCR systems reports substantial performance disparities between commercial and open-source tools, so the same manual can behave differently under a different engine.

The errors that matter on a manual

A 2026 taxonomy of OCR errors sorts them into three levels.

Characters. Insertions, deletions, substitutions, and transpositions. The authors cite look-alike pairs such as l and 1, 5 and S, and 0 and O, with an example in which blue becomes b1ue. In running prose a reader corrects that without noticing. In an identifier there is nothing to correct it against, and the result can be a valid-looking number for a different part.

Words. One word is split in two, as when greenhouse becomes green house, or two words are merged into one.

Columns. A page with several columns is read straight across instead of down each column, which interleaves text from different sections. On a procedure, that can put a step beside a warning that belongs to another step.

Two further limits come from the Tesseract documentation itself. The quality guide says the engine has trouble with text and data in tables unless custom segmentation or layout analysis is added. The FAQ says handwriting recognition is possible but will not work very well, because the engine is designed for printed text. Tables, stamps, and handwritten annotations are ordinary features of a manual that has been in service for decades.

What an accuracy score hides

Accuracy is scored against a reference transcription. A study of post-OCR correction defines character error rate (CER) as the number of character substitutions, deletions, and insertions divided by the number of characters in the reference. Word error rate (WER) applies the same calculation to words. Lower is better. Five character edits against a 500-character reference give a CER of 0.01.

Both scores are averages. They count every edit alike, so a wrong digit in a torque value and a wrong letter in a common word weigh the same. A low average says nothing about where the remaining errors sit, and on a manual the place they sit is the whole question.

Keep the confidence, and keep the location

Tesseract's output carries a confidence value for each recognized word, along with the word's position on the page. In the documentation's own sample, the first word of the first line scores about 96.

If that number is thrown away after OCR, every later stage treats every word as equally reliable. Search, translation, and question answering then inherit the misread characters with no mark showing which words to doubt.

Research supports keeping it. The same comparison study examined how confidence scores relate to error rates across several systems and reports that feeding the scores into an error-detection model improves the detection of OCR errors. The gains were largest for engines whose confidence is well calibrated, meaning the score tracks how often the word is right. A confidence score helps most when it is both kept and trustworthy.

The location matters for the same reason. With the word's place on the page stored beside its score, a person can open the original image and read the doubtful word directly.

What to ask

• At what resolution were these manuals scanned, and were the pages straightened and cropped first?
• Was accuracy measured on these manuals, or on some other test set?
• Does the reported error rate separate numbers and identifiers from running text?
• Does every extracted word keep its confidence score and its position on the page?
• Can a reader open the original image for any word in one step?

Azcuna's TRACE is one offline viewer for S1000D, MIL-STD-3001 XML, and PDF technical manuals. It keeps the source location and the markings of everything it shows and never modifies the source. It was proposed to NAVAIR under Navy SBIR Phase I topic DON26BZ05-NV078, and the proposal was submitted on September 23, 2026.

Sources

1. 1Improving the quality of the output · Tesseract OCR project (tessdoc)
2. 2FAQ (tessdoc | Tesseract documentation) · Tesseract OCR project (tessdoc)
3. 3Revise: A Framework for Revising OCRed text in Practical Information Systems with Data Contamination Strategy · arXiv (Shim, Hong, and Lim)
4. 4RoundTripOCR: A Data Generation Technique for Enhancing Post-OCR Error Correction in Low-Resource Devanagari Languages · arXiv (Kashid and Bhattacharyya)
5. 5Confidence-Aware Document OCR Error Detection · arXiv (Hemmer et al.)