There are three different ways to get content out of a PDF, and picking the wrong one is the most common reason people think a tool failed. We tested the middle one, OCR, properly: we made scans of a page of known text at different resolutions and with typical problems like tilt and blur, ran them through our PDF OCR tool, and counted the mistakes. The findings are specific: some problems barely matter and one is a trap.
Three ways to get content out of a PDF
| You have… | Use | What it does |
|---|---|---|
| A PDF where you can select text | PDF to Text | Reads the stored text layer. Exact, instant, no recognition involved. |
| A scan or photo saved as a PDF | PDF OCR | Draws each page and recognises the letters. Accuracy depends on the scan. |
| Pictures you want out of the PDF | PDF Image Extractor | Pulls the original embedded images at the size they are stored, not a picture of the page. |
The practical rule: if you can highlight a word with the cursor, the PDF already has real text, and OCR would only add errors. Use it only when the page is a picture of text. All three run in your browser, so the files are not uploaded.
How we tested OCR
We rendered one A4 page of English text (87 words, 11-point type, including numbers) at 300 DPI, then made copies at lower resolutions and with common scanning faults, saved each as a one-page scanned PDF, and ran each through the real PDF OCR tool. We compared the output with the original text and report the word error rate: wrong, missing or extra words as a share of the 87. Alongside it is the confidence score the tool itself shows.
What affected accuracy
| Scan condition | Words wrong | Characters correct | Tool's confidence |
|---|---|---|---|
| 300 DPI | 0% | 100% | 95% |
| 200 DPI | 0% | 100% | 95% |
| 150 DPI | 0% | 100% | 95% |
| 100 DPI | 2.3% | 99.6% | 95% |
| 72 DPI | 2.3% | 99.4% | 93% |
| 50 DPI | 37.9% | 90.2% | 62% |
| 150 DPI, 2° tilt | 2.3% | 99.8% | 95% |
| 150 DPI, 5° tilt | 17.2% | 96.6% | 85% |
| 150 DPI, 10° tilt | 54.0% | 86.7% | 75% |
| 150 DPI, blurred | 6.9% | 98.7% | 92% |
| 100 DPI, small 8-point text | 5.7% | 99.1% | 93% |
| 150 DPI, faint grey text on grey | 0% | 100% | 95% |
| 150 DPI, heavy noise | 0% | 100% | 95% |
| 150 DPI, JPEG quality 40 | 0% | 100% | 95% |
What the results mean
- Resolution has a ceiling. The OCR tool draws each page at twice the PDF's base size (about 144 DPI), so scanning at 300 or 600 DPI gives no gain here and makes the file larger. 150 DPI was perfect for 11-point text and 100 DPI nearly so. At 50 DPI, 11-point text is only about 8 pixels tall and more than a third of the words were wrong.
- Tilt is the trap. A page photographed or scanned at a slight angle reads worse than one that is noisy, faint or heavily compressed. 2° was fine, 5° cost 17% of the words and 10° cost over half. Straighten the page when you scan or photograph it, or crop and rotate it before making the PDF. Our Rotate PDF turns pages in quarter turns only, so it cannot fix a slight tilt.
- Compression and noise were harmless in this test. JPEG quality 40 and heavy noise changed nothing. That is good news for shrinking a scan before OCR, but our text was large and sharp, and small or blurred type will be less forgiving: blurring alone cost 7%, and 8-point text at 100 DPI cost 5.7%.
- Contrast was not the problem we expected. Faint grey text on grey paper read perfectly in this test.
Getting images out of a PDF
The PDF Image Extractor reads the picture objects stored in the file, so you get each image at its stored resolution, for example the original 2000-pixel photo placed small on a brochure page. The alternative, turning a page into an image with PDF to JPG, gives you a picture of the whole page at whatever resolution you pick. Use extraction when you want the picture itself and page conversion when you want the page.
Step by step
- Open the PDF and try to select a word. If you can, use PDF to Text and stop.
- If it is a scan, check that the page is straight and sharp. Re-scan at 150 to 200 DPI if it is not.
- Open PDF OCR and add the file. It starts automatically and shows progress per page.
- Read the confidence score, then use Copy all text or Download as .txt.
- Proofread numbers, names and dates in particular, since a single wrong digit is the costliest error.
- To edit the text, paste it into a document. The OCR tool produces text, not a searchable PDF.
What we did not test
We tested one page of clean English text in one font, and each condition once. Real documents have columns, tables, stamps, handwriting, other languages and uneven lighting, all of which lower accuracy and which we did not measure. The recognition engine runs in your browser and downloads its data the first time you use it, so that first run needs an internet connection. Treat these numbers as a guide to which problems matter, not a promise of any accuracy on your document.
