You can handle this directly in your browser with our free pdf to images tool — nothing gets uploaded.
Why Your Scanned PDF Has Garbled Text (and How to Fix OCR)
Scanned PDFs that look perfect in the browser can still produce nonsense when you copy, search, or extract text. The culprit is usually not the PDF viewer but the OCR layer beneath the images. When that layer fails, the visible document is intact while the invisible text is corrupted, which explains why the same file can open normally yet return characters like “Ϩ” or “â” instead of words. This article focuses on diagnosing why OCR produces garbled output and how to rebuild the text layer with the right rescanning parameters and verification workflow.
Understanding garbled text is a two-layer problem. The image layer stores the page as a picture, while the text layer stores the characters a reader or search engine can access. OCR sits between them, converting pixels into Unicode code points. When the conversion goes wrong, the text layer still exists, but its content is wrong. That is why simply re-running a PDF editor usually does not help. You must identify which layer is failing and whether the input image itself gives OCR enough information to work.
Garbled output has recognizable patterns. Random Latin-looking symbols often mean the source was a non-Latin script, a scanned image of a document that uses a special font, or an encoding mismatch. Diacritics turning into question marks or boxes usually points to a font that lacks the glyphs needed. Spaced-out or broken words suggest the scanner captured the page at low resolution or with motion blur. Complete gibberish, such as binary-like characters, often means the OCR engine processed an image that was not actually readable, such as an encrypted, inverted, or heavily compressed scan.
A useful first test is to check whether the PDF has any selectable text at all. If you highlight nothing, the file is a pure image and OCR was never run. If you can select text but it is wrong, the OCR layer exists and is corrupted. If text selects cleanly and matches the image, the issue is downstream, such as a search tool misreading the encoding.
Low resolution is the most common cause. Most consumer scanners default to 150 DPI, which is enough for casual viewing but too coarse for small print, fine tables, or handwriting. OCR engines need enough pixel density to distinguish similar shapes, especially I, l, and 1 or O and 0. Below roughly 200 DPI, many engines start guessing and produce phonetic or random characters.
Skew and perspective distortion is the second major cause. Even a two-degree tilt can make horizontal baselines inconsistent, causing OCR to merge or split lines. This is frequent with phone photos of documents, which are often captured at an angle. The image looks acceptable to a human but the text detection algorithm is confused by the uneven spacing.
Handwriting and stylized fonts are a third category. Standard OCR is trained on typed or printed text. Cursive, artistic fonts, and dense handwritten notes lack the consistent stroke patterns that models rely on. The result is not a small number of errors but wholesale garbling, because the engine cannot map the shapes to standard letters.
Non-Latin and mixed-script content is a fourth cause. A document that mixes Cyrillic, CJK, and Latin characters in a single page will fail if the OCR engine is configured for one script only. The engine may force every glyph into its target alphabet, producing transliteration-like output that is completely wrong. This is one of the main reasons users see “garbled” text in what they believe is a simple English scan.
Missing or embedded fonts matter more than most people assume. When a scanned PDF was created by converting a digital file that used an embedded Type 1 or CJK font, the text layer may already be broken before OCR ever touches it. In that case, OCR is not fixing the problem; it is inheriting it.
Knowing which layer is at fault decides the entire repair path. An image layer problem means the pixels are wrong: blurry, skewed, low contrast, or missing content. A text layer problem means the pixels are fine but the hidden characters are wrong or absent.
Start by viewing the PDF at 200 percent zoom. If the words are sharp, the image layer is healthy and the fault is in the text layer. If the words are fuzzy, the image itself is the problem and no OCR pass will save it. Next, try copying a clean block of text. If it returns correct characters, the text layer is good. If it returns boxes or symbols, the text layer is broken even if the image is fine. Finally, use a PDF tool that displays the text layer separately, such as a viewer that can hide images or a tool that exports the text stream. If the exported text is empty, OCR was never applied. If it is full of nonsense, OCR was applied badly.
This three-step check saves time because it stops you from re-scanning a file that only needs its text layer regenerated, or from running more OCR on a file whose image is fundamentally unreadable.
When the image layer is the problem, rescanning is the correct move. The following parameters are the starting points that consistently improve recognition rates for mixed document types.
Set the resolution to 300 DPI minimum. For typed text with large fonts, 200 DPI may suffice, but 300 DPI is the safe default and costs only a small amount of storage. For fine tables, small footnotes, or anything under 9 point, move to 400 to 600 DPI. The key is that the smallest characters in the document should be at least 8 pixels tall after scanning.
Scan in grayscale rather than color for most text documents. Grayscale preserves enough contrast for OCR while producing smaller files that are faster to process. Use color only when the document contains colored text, stamps, or highlights that you need to preserve visually. For pure OCR purposes, color adds no accuracy and increases processing time.
Correct skew before OCR. Most OCR suites have an automatic deskew function, but it works best when the initial tilt is under about five degrees. If the scan is tilted more, rotate the page manually first. A consistently skewed set of pages is a leading cause of line-merge errors.
Enable contrast enhancement or adaptive thresholding if the scan is faded or the paper is yellowed. This step converts a gray-toned image into a clean black-and-white or gray image that OCR engines read more reliably. Overdoing it, however, can erase faint text, so preview the effect before committing.
Choose the correct language and script setting. If the document is English, select English. If it is multilingual, either run OCR twice with different language profiles or use a profile that covers the dominant script. This single setting prevents the majority of “garbled” output in non-Latin documents.
Select a 600 DPI or higher source if you are rescanning from a phone. Phone cameras produce large images, but the effective optical resolution is often far below 600 DPI unless the lens is in focus and the subject is parallel. A tripod or a flat surface and adequate lighting matter more than the megapixel count.
Even with good parameters, OCR is not perfect. A structured proofing pass catches the remaining errors before the document is relied on.
First, read a random sample of pages, not just the first page. Errors cluster in areas with tables, headers, and footnotes, so sample those. Second, copy a few known phrases and verify they match character for character, including special punctuation and currency symbols. Third, use a spell checker that supports the target language, but treat it as a filter, not a verdict. OCR errors are often real words that are just wrong, such as “recieve” instead of “receive,” and spell checkers will not flag them. Fourth, cross-check any numbers, dates, and identifiers against the source, because these are the fields where a single wrong digit is the most damaging. Finally, save the corrected version as a separate file and keep the OCR output for reference, so you can audit changes.
For large documents, automate the easy checks. Compare the character count of the extracted text to the expected range for the page, run a script that flags lines with unusually high symbol density, and spot-check pages that fail those flags. This keeps the human proofing effort focused on the pages that actually have problems.
Some inputs are beyond OCR no matter how carefully you rescan. The most common are heavily watermarked or overlapping text, where two fonts occupy the same space. OCR will pick one and ignore the other, so the missing half is lost. Faded or coffee-stained pages where the ink contrast is below the threshold of the sensor produce random characters because there is no reliable signal. Handwritten mathematical formulas, dense diagrams with embedded labels, and cursive notes from before the 1990s are routinely misread because training data for those styles is sparse.
Encrypted or redacted scans are a separate category. If the original digital file was encrypted and then printed and scanned, the content is present but the OCR engine sees only the printed page. If the page was redacted with a black box before printing, the text under the box is gone permanently and OCR cannot reconstruct it. Likewise, a scan of a scan, where the original is already a low-resolution photocopy, compounds the loss of detail until no engine can recover the characters.
When a document falls into these categories, the honest answer is that OCR will not fix it. The only reliable remedy is to find the original digital source or to retype the critical fields by hand. Attempting to force an OCR pass on such a file gives false confidence and produces a text layer that looks plausible but is wrong in detail, which is worse than having no text layer at all.
Why does my scanned PDF show random symbols instead of letters? That is usually an encoding or script mismatch. The OCR engine forced the glyphs into the wrong character set, so the visible image is fine but the hidden text is corrupted. Re-run OCR with the correct language profile and regenerate the text layer.
Can I fix garbled text without rescanning the document? Yes, if the image layer is sharp. You can delete the existing text layer and run OCR again with better parameters. If the image is blurry or skewed, rescanning is necessary because no amount of OCR tuning can recover detail that the pixels do not contain.
What DPI should I use for a typical business document? 300 DPI in grayscale is the standard. Go to 400 or 600 DPI only for small type, fine tables, or detailed images. Higher DPI does not improve accuracy on large, clear text and only increases file size.
How do I know if the problem is the image or the text layer? Zoom to 200 percent and judge the visual sharpness. Then copy a block of text. If the image is sharp but the copied text is wrong, the text layer is broken. If the image itself is fuzzy, the image layer is the fault and you must rescan.
Related reading: annual review scan reconcile paper records, batch ocr scanned pdfs, car repair records scan before selling.