What OCR can and cannot read reliably

OCR is useful because a scanned page is only a picture until something reads the letters inside it. After recognition, you can search, copy, index, summarise or rebuild the text layer. That usefulness can make OCR sound simpler than it is: put in a scan, get back the words.

The reality is a pipeline. Each stage depends on the previous one. If the page image is poor, or the layout is misunderstood, the recogniser may produce fluent text that is still wrong. This post is about calibrating expectations. For the step-by-step workflow, use the guide to make a scanned PDF searchable with OCR. For handwriting, use the separate guide to read handwritten notes into searchable text.

OCR starts before recognition

Recognition is the final visible step, but it is not where most failures begin. A typical OCR engine first prepares the page image. Binarisation tries to separate foreground text from background paper. Deskew straightens a page that was scanned at a slight angle. Noise removal tries to ignore specks, dust and compression blocks.

Then the engine has to understand the document structure. It identifies text regions, lines, words and sometimes individual characters. Only after that does the recogniser decide whether a mark is an O, a 0, an I, an l, a 5, or an S.

A failure at an early stage poisons everything after it. If binarisation turns a faint stroke into background, the recogniser never sees that stroke. If deskew leaves text tilted, line detection may cut through letters. If segmentation splits rn into something that looks like m, no dictionary can reliably recover the original word in every case.

Layout analysis is the underrated step

People often judge OCR by whether individual letters were recognised. For real documents, layout analysis is just as important. It decides what order the text should be read in, where one column ends and another begins, whether a caption belongs to an image, and whether a table should remain a table or collapse into a stream of words.

A two-column report can become nonsense if the engine reads across both columns line by line. A bank statement can lose meaning if amounts are separated from dates and descriptions. A form can look accurate at the character level while being useless because labels and values no longer sit together.

Use the OCR PDF tool when the goal is a searchable text layer, the scan document tool when the image still needs capture and perspective correction, and the text extraction tool when the PDF already contains real text and does not need recognition.

Inputs that reliably defeat OCR

Some pages are simply bad evidence. OCR cannot read what the image does not contain clearly enough. The most common problem inputs are:

• Low-resolution scans where letters are only a few pixels high.
• JPEG artefacts around strokes, especially after repeated sharing.
• Skewed pages, camera blur, shadows and curved book pages.
• Background patterns, stamps, seals or handwriting crossing printed text.
• Faint thermal receipts, carbon copies and low-contrast photocopies.
• Decorative, unusual or heavily condensed typefaces.
• Mixed scripts or languages without the right language model enabled.

You can sometimes improve the page before recognition: rescan at a higher resolution, flatten a curled page, improve lighting, crop away the desk, or capture from directly above. But there is a limit. A recogniser cannot infer a digit that was never captured with enough detail.

Dictionaries help and also mislead

Modern OCR often uses language models, dictionaries or contextual correction. That helps with ordinary prose. If the raw character pass reads rec0gnition, a language-aware pass can prefer recognition. It can also distinguish common letter shapes by looking at surrounding words.

The risk is that context can turn uncertainty into a plausible wrong answer. Obvious garbage is easy to spot. A real word substituted for another real word is more dangerous, especially in contracts, invoices, prescriptions and identity documents. Public and PubIic may look similar. B8 and 88 can be hard to separate in a poor scan. A dictionary may prefer a common surname over the rare surname actually printed.

That is why OCR output should be treated as a draft. It is excellent for making a document searchable and reducing retyping. It is not a witness you can trust without checking when the exact characters matter.

Percentages need a denominator

OCR accuracy figures are often quoted as percentages, but a bare percentage is not meaningful. Is it measured per character, per word, per field, or per page? Was the material clean office printing, aged newsprint, receipts, forms, tables, or multilingual documents? Were punctuation, line breaks and reading order counted?

A 99 percent character accuracy rate sounds high. On a page with 3,000 characters, it still allows about 30 wrong characters. If the errors land in body prose, you may not care. If one lands in an account number, passport number, medicine dose, date, GSTIN, or email address, the cost is disproportionate.

Word accuracy can look worse because one wrong character makes the whole word wrong. Field accuracy can look worse again because one wrong digit makes the field unusable. Always ask what the number measures and on what kind of material.

What to proofread by hand

The practical rule is simple: proofread anything that identifies a person, amount, place, obligation or deadline. Names, numbers and identifiers are exactly where a confident substitution hurts most. Check totals, dates, invoice numbers, policy numbers, registration numbers, tax IDs, account numbers, addresses and clause references against the page image.

For long documents, you do not have to proofread every paragraph with equal care. Searchability may be enough for narrative text. But fields that drive a payment, filing, shipment, medical decision or legal obligation deserve a manual compare. OCR saves time; it does not remove responsibility for the final read.

A short OCR checklist

• Start with the cleanest image you can capture, not the fastest one.
• Check whether the page is skewed, shadowed, curved or compressed.
• Inspect columns, tables and forms for reading-order errors.
• Treat dictionary-corrected words as suggestions, not proof.
• Proofread names, numbers and identifiers against the original page.