OCR Output Errors: What to Check

When OCR (optical character recognition) output contains errors, the first things to check are the source image quality, the language and font settings, and the specific characters that were misread. Fixing these three areas resolves the vast majority of accuracy problems before you need to consider more advanced correction tools.

Formula

The calculator above gives a rough idea of how much resolution you have relative to the commonly recommended 300 DPI for printed text. The formula is:

Effective quality (%) = (Image DPI ÷ 300) × 100

Where:

• Image DPI – the resolution of the scanned image or photo (dots per inch)
• 300 – the typical minimum DPI for clean OCR of most printed documents
• Rotation angle – included as a reminder that skewed images cause errors; the calculator itself does not correct rotation

This is a heuristic, not a guarantee: a 300 DPI image of a blurry photo will still OCR poorly, while a 150 DPI image of crisp, large text can work fine.

Worked Example

Using the default values: you have an image at 300 DPI and 0 degrees rotation.

1. Divide the DPI by 300: 300 ÷ 300 = 1.0
2. Multiply by 100: 1.0 × 100 = 100%

The calculator shows 100% effective quality, meaning you are at the recommended threshold. If you had a 150 DPI image, it would show 50%, which is a strong hint that resolution is contributing to errors.

Why OCR Errors Happen

OCR software works by recognizing shapes of letters and numbers. It does not “read� like a human; it matches patterns. Any deviation from the expected pattern – a smudge, a fold, a low-contrast scan, an unusual font – can cause a mismatch. The most common errors are:

• Confusable characters: lowercase l vs. number 1, capital O vs. zero, rn vs. m, cl vs. d.
• Missing or extra spaces: especially in justified text or narrow columns.
• Garbled punctuation: apostrophes, quotes, and hyphens get replaced with random symbols.
• Whole-word substitutions: when the software guesses a word that looks similar but is wrong (e.g., “rheâ€� instead of “theâ€�).

Knowing these patterns helps you know what to look for when proofreading.

Image Quality Checks

Before touching any settings, inspect the image itself. The most common fixable issues are:

• Resolution: Below 200 DPI, small text becomes a jagged mess. Re-scan at 300 DPI or higher.
• Skew: A page rotated even 2–3 degrees makes lines of text curve out of the recognition grid. Straighten it with any image editor or the OCR tool’s auto-deskew.
• Contrast: Faded ink or dark backgrounds confuse the threshold between letter and paper. Increase contrast or use a “grayscaleâ€� or “black and whiteâ€� mode.
• Noise and artifacts: Dust, speckles, or compression artifacts from a JPEG can create false “letters.â€� Use TIFF or PNG for scans.
• Lighting (for photos): Shadows across the page create gradients that break character shapes. Use even, diffuse lighting.

If you are working with a photo of a document rather than a scan, these issues are far more likely.

Software Settings to Verify

Most OCR tools have settings that dramatically affect accuracy. Check these before reprocessing:

• Language: If the document is in English but the OCR is set to French, you will get accented characters and wrong word breaks. Set the exact language (and any regional variants, like Portuguese vs. Brazilian Portuguese).
• Document type: Is it set to “scanned document,â€� “book,â€� “receipt,â€� or “photoâ€�? Each uses a different recognition model.
• Text orientation: Some tools need to know if the text is landscape, portrait, or mixed.
• Dictionary/lexicon: Enabling a dictionary helps the engine choose real words over gibberish, but it can also “correctâ€� proper nouns or technical terms. For scientific papers, turn it off.
• Character whitelist: If you only need digits, restrict the OCR to numbers. This eliminates letter/number confusion.

These settings are often buried in a “Preferences� or “Options� menu, not on the main screen.

Common Error Patterns and Fixes

Even with perfect settings, errors slip through. Here are the most frequent patterns and how to fix them:

• “lâ€� vs. “1â€� vs. “Iâ€�: In fonts without serifs, these look identical. Check the context: if it is in a number sequence, it is likely 1; if in a word, it is l or I.
• “Oâ€� vs. “0â€�: Same issue. Many OCR tools have a “zero with slashâ€� option for numeric fields.
• “rnâ€� vs. “mâ€�: This happens when the two letters are too close. Increase DPI or use a font with clearer spacing.
• Missing diacritics: Accents on é, ñ, ü get dropped or replaced. Ensure the language pack includes the right character set.
• Hyphenation artifacts: Words split across lines (e.g., “tele-â€� + “phoneâ€�) get a hyphen that should be removed. Some tools have a “dehyphenateâ€� option.

If you see the same error repeatedly, use the tool’s “replace� feature or a post-processing script.

Post-Processing and Proofreading

OCR output should never be trusted as final. A practical workflow:

1. Run the OCR and export to a text editor or word processor.
2. Use spell-check to catch obvious garbles, but be aware it will miss correctly-spelled wrong words (e.g., “form� instead of “from�).
3. Search for common confusables: “0� in words, “1� in words, “rn� in words.
4. If the document is long, sample-read a few paragraphs from different pages rather than the whole thing.
5. For critical documents (legal, medical), consider a second OCR engine and compare outputs.

There are also automated tools that use language models to correct OCR errors by context, but they can introduce new mistakes. Always keep the original image for verification.

When to Use a Different OCR Engine

No OCR engine is perfect, and they have different strengths. Tesseract is free and open-source, but it struggles with complex layouts. ABBYY FineReader handles tables and multi-column pages well, but it is commercial. Google Cloud Vision and Azure Cognitive Services are good for photos of documents because they use neural networks. If you are getting poor results with one engine, try another; the error patterns will likely be different, and you can cross-check.

For handwritten text, specialized handwriting OCR (ICR) is needed; standard OCR will fail almost completely.

Frequently Asked Questions

Why does OCR turn “0� into “O� so often?

Because in most fonts, the digit zero and the capital letter O are nearly identical. The software uses context to decide, but in a string of digits (like a phone number) it has no context. Fix by using a font with a slashed zero, or post-replace “O� with “0� in numeric fields.

Can I fix OCR errors automatically?

Yes, partially. Spell-checkers catch many, but they miss correctly-spelled wrong words. Context-aware tools (like those using language models) can correct more, but they can also “correct� a correct word into a wrong one. Always review automatic corrections.

What is the best DPI for OCR?

300 DPI is the standard recommendation for printed text. For small fonts (below 8 point), use 400–600 DPI. For handwriting, 300 DPI is usually enough, but the writing must be clear. Higher DPI does not always help if the image is blurry.

Why does OCR struggle with tables and columns?

OCR engines are trained on linear text. Tables have lines, varying column widths, and text that flows in boxes, which breaks the linear model. Use a tool with table detection, or pre-crop each column and run OCR separately.

Should I use PDF or image for OCR?

If the PDF is already digital (born-digital), it has a text layer and OCR is unnecessary. If it is a scanned PDF, you can either OCR the PDF directly or export pages as images and OCR those. Direct PDF OCR is convenient, but image-based OCR often gives more control over preprocessing.

Does OCR work on handwriting?

Standard OCR is designed for printed characters. Handwriting requires ICR (Intelligent Character Recognition), which is less accurate and often needs training on a specific writer’s style. For a single page of neat handwriting, you might get 70–80% accuracy, but printed text can reach 99%+ under good conditions.

Related

• Second-Level Domain: What It Means Before the TLD
• Google Play Music to YouTube Music Transfer Tool
• IBM Bob vs GitHub Copilot Business: On-Premises Data Control

in t | Image resolution (DPI) | dpi | 300 | 72 | 1200
in r | Rotation angle | degrees | 0 | -180 | 180
out quality | Effective scan quality | % | min(100, (t / 300) * 100) | 0