Making a scanned PDF searchable without trusting every recognised word

OCR adds machine-readable text to a scanned page. It can make a document much easier to search, but the recognised text is an interpretation of the image and needs its own quality checks.

Recognising words is not recovering the original document

A scan records appearance. OCR attempts to infer characters and their positions from that appearance. It does not retrieve the original word-processing file, spreadsheet formulas or editing history. Even a good recognition result can have the wrong reading order, confused punctuation or an incorrect digit. The page image and the text layer should therefore be treated as related but distinct representations.

This matters when you search or copy. A page may visibly say “1,250” while its text layer says “I,250.” A search for the correct amount may fail, or a copied value may be wrong despite a perfectly readable image. The error becomes more serious if that text is then fed into a summary, translation, spreadsheet or document-chat tool without review.

Use OCR to make access easier, then verify the parts that matter for your purpose. A personal reading collection might tolerate occasional punctuation errors. An invoice total, identifier or safety instruction deserves character-by-character comparison. There is no single accuracy percentage that makes every downstream use safe.

Good capture is usually better than aggressive repair

Start with a sharp image. Keep the page flat, the camera steady and the lighting even. Avoid shadows over text, reflected light on glossy paper and perspective distortion from photographing at an angle. If a page is curved near a book’s binding, important letters may be squeezed or obscured before the recognition engine sees them.

Increasing resolution helps only when it captures real detail. Enlarging a blurry photograph creates more pixels without restoring the missing strokes. Tesseract’s documentation discusses resolution, skew, noise and page segmentation as important input factors; it notes that recognition generally benefits from sufficiently detailed images. See Tesseract’s image-quality guidance for the technical background.

Before processing a long document, inspect its weakest pages. A clear title page tells you little about faint carbon copies, small footnotes or a densely printed appendix. If you can rescan a bad page now, that may save more time than correcting hundreds of recognition mistakes later. Keep the original capture even when you make a cleaned working copy.

Choose the language deliberately

Recognition models encode expectations about characters and word patterns. Selecting the wrong language can increase errors in accented words and common letter combinations. SandboxPDF’s OCR tool currently offers English, French, German and Spanish. These choices are separate from the languages supported by its summary or translation models; one feature’s language list does not imply support in another.

For mixed-language pages, inspect the less dominant language particularly carefully. Names, addresses, product codes and foreign quotations may not follow the model’s usual vocabulary. A plausible-looking correction can be more dangerous than obvious gibberish because it is easy to overlook. Never let a recognised spelling silently replace an authoritative name or identifier in a final record.

If your required language is not supported, use a suitable recognition tool rather than selecting the nearest-looking option and assuming the result is reliable. The image can remain useful even when the text extraction is poor. Keep those two quality judgments separate.

Test a representative page first

Choose a page containing the document’s actual challenges: columns, a small table, a date, a monetary amount and a few accented words. Run OCR PDF with the relevant language. On first use, the browser may download a recognition model. That download does not mean the document is being uploaded; recognition runs locally after the required assets are available.

Read the output alongside the page image. Test a search for a distinctive phrase, copy a paragraph into a plain-text editor and inspect the order of the words. Check whether line endings introduce unwanted breaks. If the page has two columns, confirm that the extracted text does not alternate between them line by line.

Only proceed to the full selection once this sample is useful. If the representative page fails, processing another hundred similar pages is unlikely to solve the underlying problem. Improve the input, change the workflow or plan a manual correction step. A short pilot is especially worthwhile on a phone, where recognition can be slower and memory more limited.

Numbers deserve a separate pass

OCR errors involving letters can sometimes be understood from context. Numbers often cannot. Confusions between zero and the letter O, one and lowercase l, or five and S can change identifiers. A missing minus sign can reverse meaning. A misplaced decimal separator can change an amount by a factor of one hundred or more.

Build a small verification list before you begin: invoice totals, dates, account references, quantities and other fields that will be reused. Compare those fields against the source image after recognition. Preserve leading zeros in identifiers. Do not allow a spreadsheet’s automatic type conversion to turn an account code into a number or a date while you are checking OCR accuracy.

For a table, validate several complete rows, including the relationship between labels and values. Correct characters placed in the wrong column are still incorrect data. Totals can provide a useful cross-check, but an arithmetically consistent table is not proof that every source value was recognised correctly.

Searchability and accessibility are different goals

A recognised text layer makes some content searchable and copyable. A fully accessible PDF also needs meaningful structure, reading order, labels and other information appropriate to the document. OCR by itself does not establish headings, table relationships or useful alternative text for diagrams. A file can be searchable and still difficult to use with assistive technology.

If accessibility is a delivery requirement, assess it separately using the relevant authoring and validation tools. Returning to an editable source document is often better than trying to infer structure from a scan. When no source exists, remediation may involve human decisions that a recognition engine cannot make reliably.

Do not replace an accessible digital PDF with a raster copy and assume OCR will restore its previous quality. The recognised words may look similar while semantic structure has been lost. Choose operations that preserve the document’s useful features whenever possible.

Plan the order of operations

Rotate sideways pages before recognition so that the text is presented in a sensible orientation. Improve an obviously poor capture before spending time on extraction. If redaction is required, create and verify the redacted visual copy first, then run OCR only on that copy. Do not attach an original text layer that contains information you intended to remove.

Be careful with compression after OCR. SandboxPDF’s image-based compression rebuilds page images and removes the searchable layer. A lossless operation is the more appropriate first attempt when you need to retain recognised text. Check the final downloaded file after the complete workflow, because an intermediate OCR result can be correct while a later transformation removes it.

If you plan to summarise or translate the document, inspect the extracted text before invoking those tools. They cannot reliably correct an OCR error simply because the resulting sentence sounds fluent. A summary of a misread amount or a translation of a missing negation may be convincing and wrong.

Work within the browser’s practical limits

Recognition uses CPU time and memory. A long, high-resolution scan can require much more working memory than its compressed file size suggests. Keep the tab open during processing, avoid starting several heavy jobs at once and use a smaller page selection if the device struggles. A desktop may be the better choice for a large batch.

Download results before closing the session. The workspace is not a cloud document library, and refreshing can discard work held in memory. Browser caches may retain model assets, but that is different from retaining your document or completed output. Store the verified result in a location you can find again.

If processing fails, record whether it failed during model initialization, page rendering or recognition. Those stages have different causes. A network problem during first-time setup does not imply that the PDF itself is corrupt, while a consistent failure on one page may indicate an input-specific issue worth isolating.

Decide whether the result is fit for its intended use

A useful OCR review asks concrete questions: Can I find a known phrase? Does a copied paragraph read in the right order? Are the important numbers exact? Are all selected pages present? Does the image still show the original evidence clearly? Can the intended recipient use the result in their own reader?

Keep a record of any manual corrections made to extracted data, especially when moving values into another system. Retain the source image so that a reviewer can trace a value back to the page. Where accuracy has legal, financial or safety consequences, use an appropriate independent review process rather than relying on a general recognition tool alone.

For problems with the order of copied text, continue with PDF reading order. For files that are too large to handle comfortably, read the scanned-PDF size guide before choosing a compression method that could discard the text layer you just created.