How to Prepare Scanned Documents for More Accurate OCR

For one image where you need text, line confidence and JSON, use Advanced Image OCR. Use PDF OCR when you need to keep a PDF searchable, or Scanned PDF to Word when you need an editable DOCX draft. None of these replaces a human review, especially for names, amounts, IDs, legal text and tables.

This walkthrough uses Advanced Image OCR because it returns TXT, structured JSON, line counts and average confidence. That makes it easier to see what was recognized and where to start checking.

Enlarge image

Self-authored English scan sample. It includes paragraphs, numbers, status words and table lines for a deliberate OCR review.

Start with the clearest source available. Keep the page upright, avoid shadows and glare, crop excessive empty margins, and preserve enough resolution for the smallest text. A clear, front-facing scan is normally more useful than a larger but blurry photo.

The English sample deliberately keeps text, numbers, status words and table lines so you can compare each part after recognition. Do not upload identity documents, contracts or private photos just to test a tool.

• Prefer the original scan over a screenshot of a screenshot.
• For phone photos, flatten the page, use even light and avoid perspective distortion.
• When a file mixes languages, choose the closest language and manually inspect the other language.
• Test one representative difficult page before processing a whole archive.

Open Advanced Image OCR above. The screenshots come from the real desktop interface; on a narrow screen the controls stack vertically, but the order stays the same.

1. 1

   Choose one image.

   Select mvtools-ocr-scan.png in the image input. Selecting a file does not start recognition.

2. 2

   Choose the closest OCR language.

   This sample uses English. The language changes the recognition model and character interpretation; inspect sections in other languages carefully.

3. 3

   Enable preprocessing when capture conditions need it.

   This sample leaves image preprocessing enabled. It can help with low contrast, uneven light or photographed pages, but cannot restore detail missing from the source.

Enlarge image

Figure 1 · Confirm the filename, choose the closest language and decide whether preprocessing is appropriate.

Click Run PaddleOCR and wait for the task to finish. Do not submit repeatedly while it is processing. When complete, read the status, language, line count and average confidence before downloading text or JSON.

1. 1

   Confirm that the task completed.

   COMPLETED means result files were produced; it does not mean every character is correct.

2. 2

   Use confidence as a sorting clue.

   This sample averages 99.7%. Check low-contrast small text, numbers, symbols and table edges first, then review the rest.

3. 3

   Download both result formats.

   TXT is convenient for reading and editing. JSON preserves per-line confidence, boxes and structured details for deeper checks.

Enlarge image

Figure 2 · An actual English run returned 18 lines from a 96 KB image with 99.7% average confidence.

Read the TXT from beginning to end, then open JSON and inspect each line’s confidence, box and polygon. Compare the source and output side by side, focusing on areas most likely to affect the real task.

The page shows a preview of this sample output. The preview is not a substitute for checking the complete files; download the scan, TXT and JSON from the example above.

Enlarge image

Figure 3 · A TXT preview is useful for a quick read, but numbers, symbols and table relationships still need the source image.

• Text: check for missing, repeated or substituted characters and unexpected spaces.
• Numbers and symbols: verify IDs, dates, decimal points, hyphens and units one by one.
• Structure: check table rows and columns, paragraph order, stamps, signatures and handwriting.
• Purpose: confirm the output meets your editing, search or archive goal; keep the original scan for important content.

Enlarge image

Self-authored English scan sample. It includes paragraphs, numbers, status words and table lines for a deliberate OCR review.

Enlarge image

Figure 3 · A TXT preview is useful for a quick read, but numbers, symbols and table relationships still need the source image.

Advanced Image OCR accepts one JPG, PNG, TIFF, BMP or WebP image up to 25 MB, with a default five-minute task timeout. PDF OCR accepts PDFs up to 100 MB and 100 pages; Scanned PDF to Word accepts up to 100 MB and 50 pages. Encrypted or damaged PDFs may fail.

These tasks upload files to the server. Files are normally retained for 30 minutes and then cleaned up. Download results promptly, do not treat a task link as cloud storage or permanent sharing, and upload only files you are authorized to process.