PDF text extraction or OCR?

A PDF page can contain text, images or both

A PDF exported from Word, InDesign or a browser commonly stores characters along with fonts and their positions. A scanner commonly creates a page-sized image instead. Some scanning software adds an invisible OCR text layer behind the image, making the document searchable even though the visible page still looks scanned.

Open the document and try to select individual words. Search for a distinctive phrase. If selection and search work, direct extraction is the appropriate first method. If only a rectangular image can be selected, the file probably needs OCR.

Why extracted text can lose columns and tables

PDF describes where objects appear on a page; it is not necessarily a stream of paragraphs like an EPUB or Word file. Two-column articles, sidebars, headers, footers and table cells can therefore be returned in an unexpected order. Font encoding can also produce missing characters when the PDF does not map glyphs cleanly to Unicode.

Use extracted TXT as a working copy, not automatic proof that meaning was preserved. Compare headings, numbers, dates and table values with the source, especially when the text will support legal, financial or technical decisions.

Use OCR only when the pixels are the source

Optical character recognition analyzes the shapes in a page image and predicts characters and words. Results depend on scan resolution, language, typeface, skew, contrast, handwriting and page damage. OCR can make an archive searchable, but confidence is not the same as correctness.

For stronger OCR, begin with a straight, evenly lit scan around 300 PPI, select the correct document language and proofread names, decimal points, punctuation and similar characters such as O and 0. Retain the page image so uncertain passages can be checked later.

How optical character recognition actually works

OCR transforms pixel patterns into characters through three main stages. First, image preprocessing cleans the scan by removing noise, adjusting contrast and normalizing brightness. Next, feature extraction identifies candidate characters by comparing shapes to known templates or machine learning models. Finally, a language model analyzes character sequences to predict the most likely words and correct errors based on context, grammar and common patterns.

Modern OCR engines use neural networks trained on millions of examples to recognize characters, but the process remains fundamentally probabilistic. A single pixel error can change a '5' to an 'S' or an 'O' to a '0'. The language model helps, but technical documents with unusual symbols, non-Latin scripts or poor scan quality still require human review.

Scan quality factors that affect OCR accuracy

OCR performance depends heavily on the quality of the original scan. A resolution of 300 PPI (pixels per inch) provides sufficient detail for most printed text, while 600 PPI may be needed for small fonts or high-quality documents. Page skew (tilt) confuses character recognition; most tools automatically correct skew up to 5 degrees, but larger angles require manual straightening. Contrast between text and background should be at least 60% to distinguish characters clearly.

Lighting should be even across the page to avoid shadows or glare. Damaged pages with folds, tears or handwritten annotations create additional challenges. Color scans can sometimes improve recognition by providing more detail than black-and-white, but many OCR tools work best with grayscale or monochrome images to reduce complexity.

Ligatures, hyphenation and broken encoding in extracted text

PDF text extraction can reveal artifacts from the original document's encoding and typesetting. Ligatures (combined character pairs like 'fi' or 'ae') may appear as single characters or separate letters depending on the font and encoding. Hyphenated words at line breaks can become 'hyphenated' or 'hyphen ated' depending on how the PDF stores the hyphenation point. Special characters like em dashes, smart quotes and mathematical symbols may be replaced with question marks or other placeholders if the font mapping to Unicode is incomplete.

Broken encoding is common in older documents or those created with non-standard fonts. Characters may appear as boxes, question marks or incorrect symbols. The UTF-8 format preserves most modern characters, but legacy encodings like Windows-1252 or MacRoman can cause display issues. Always review extracted text for these artifacts, especially in professional or academic work where character accuracy matters.

Which PDF viewers expose text export capabilities

Most modern PDF viewers provide text export, but the quality and method vary. Adobe Acrobat Reader offers 'Save as Other' → 'Text' with options for plain text or accessible text. Preview on macOS includes 'Export as Text' in the File menu. Browser-based PDF viewers typically allow text selection and copy, but may not provide bulk export. Some specialized viewers like Foxit Reader and PDF-XChange Editor offer advanced text extraction with formatting preservation.

Browser print-to-PDF workflows generally do not preserve text layers, creating image-based PDFs that require OCR. When choosing a viewer for text extraction, consider whether it maintains the original reading order, handles columns and tables correctly, and supports the languages in your document. Always test with a sample page before processing the entire document.

Review workflow for numbers, dates and names

When extracting text for important documents, establish a systematic review process. For numbers, check decimal points, currency symbols and mathematical operators. Dates should be verified for correct format and century (especially with two-digit years). Names require special attention to diacritics, middle names and suffixes. Automated extraction often misreads similar characters like 'l' and '1', 'O' and '0', or 'Z' and '2'.

Create a checklist of critical information types and verify each against the original PDF. Use search functions to find specific patterns like phone numbers, email addresses or identification numbers. For documents with repeated information (like invoices or forms), cross-reference extracted values across multiple instances. This systematic approach catches most recognition errors before the text is used for decision-making.

When manual transcription beats both extraction and OCR

Manual transcription remains the most reliable method for short, critical passages or when accuracy is non-negotiable. Legal documents, financial statements, medical records and technical specifications often contain information where a single character error could have serious consequences. For these cases, the time spent manually verifying text is justified by the reduced risk of errors.

Manual transcription is also preferable for highly sensitive material where OCR confidence scores are low, or when the document contains complex formatting that neither direct extraction nor OCR handles well. While slower than automated methods, manual review provides human judgment that can catch context-dependent errors and ensure the final text matches the intended meaning rather than just the visible characters.

A private direct-extraction workflow

Choose the PDF in File Preflight and let the browser read each page's existing text layer. Review the on-screen preview, correct obvious reading-order problems if appropriate, and download the UTF-8 TXT result. The selected PDF and extracted text remain in the browser tab until you close or replace them.

• Confirm that you are authorized to reuse the document text.
• Search the output for missing or replacement characters.
• Check page transitions, columns and repeated headers.
• Use an OCR workflow separately when no selectable text is found.

A worked extraction review session

Take a 24-page report extracted from a born-digital PDF. Eight spot-checks catch nearly every real problem: the totals in any table, every date, every person and company name, hyphenated words at line ends, ligature pairs such as fi and fl, footnote markers, repeated headers that belong to the page not the text, and the transitions where one page becomes the next.

In practice the numbers and names are where extraction actually fails: a decimal point that migrated, or two swapped table columns, changes meaning while still reading fluently. Comparing those eight categories against the source takes about fifteen minutes on a 24-page document, and anything that passes them is usually fit for reuse.

Scan quality arithmetic

A US Letter page scanned at 300 PPI produces 2550 x 3300 pixels — 8.5 times 300 and 11 times 300. The same page at 72 PPI produces 612 x 792 pixels, and at that density the average printed character is only a dozen pixels tall. OCR accuracy collapses well before the image looks obviously bad on screen, because the shapes lose the detail that distinguishes an 8 from a 6.

Three hundred PPI is the practical floor for text OCR, and 400 PPI buys a margin for small type and photocopy noise. Straightening a skewed scan before OCR helps as much as adding resolution: recognition drops measurably even at small angles, and deskewing is a one-click operation in most scanning tools.

• Letter at 300 PPI = 2550 x 3300 px; at 72 PPI only 612 x 792 px.
• 300 PPI is the floor for reliable text OCR; 400 PPI for small type.
• Deskew before OCR — angle hurts recognition as much as low resolution.

Languages, and documents that mix them

OCR accuracy depends on the recognition language being set correctly, because the language model resolves ambiguous shapes: the same glyph cluster can be a valid word in one language and nonsense in another. A document that switches languages mid-page — a contract with an English body and a French exhibit — should be processed with both language packs enabled where the software allows it, and the language boundary treated as a spot-check zone.

Scripts matter more than languages. Latin-script OCR is mature; accented characters add occasional errors, while mixed-script documents can interleave two character sets in one line and confuse segmentation. For any document where a single wrong character changes meaning — part numbers, dosages, account figures — plan a manual pass over exactly those fields regardless of the reported confidence.