How do I extract text from low-quality scans using OCR?

Extracting text from low-quality scans with OCR works best when you improve the image first and then tune the OCR engine.

1) Clean up the scan before OCR

Preprocessing often makes the biggest difference.

• Deskew the page if it’s tilted.
• Increase contrast and convert to grayscale.
• Denoise speckles and scanner artifacts.
• Binarize carefully: adaptive thresholding usually beats a simple black/white cutoff on uneven scans.
• Sharpen lightly if text is blurry.
• Upscale small text 2×–4× before OCR.
• Crop borders and remove background shadows.
• If the page has curved pages or book gutter distortion, use dewarping if available.

2) Use a strong OCR engine

Good options:

• Tesseract: free, widely used, decent with good preprocessing.
• PaddleOCR: often better on noisy or complex layouts.
• ABBYY FineReader: commercial, often excellent on poor scans.
• Google Cloud Vision / Azure OCR / AWS Textract: strong cloud APIs.

For low-quality scans, modern deep-learning OCR tools often outperform classic OCR.

3) Set OCR options for the document

• Choose the correct language(s).
• If the page has a known layout, use the right page segmentation mode.
• For single text blocks, line OCR can be better than full-page OCR.
• If the scan includes multiple columns, detect and split columns first.

4) Post-process the OCR output

• Run spell check and language correction.
• Use domain-specific dictionaries for names, medical/legal/technical terms.
• Fix common errors like:
  • l vs 1
  • O vs 0
  • rn vs m
  • broken hyphenation across lines
• If accuracy matters a lot, compare OCR outputs from two engines and merge results.

5) Practical workflow

A good pipeline is:

1. Scan at 300 DPI or higher if possible.
2. Preprocess image.
3. Run OCR.
4. Clean and correct text.
5. Manually review low-confidence sections.

6) Example with Tesseract + preprocessing

Using Python and OpenCV:

import cv2
import pytesseract

img = cv2.imread("scan.jpg")

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
gray = cv2.fastNlMeansDenoising(gray, h=20)
gray = cv2.resize(gray, None, fx=2, fy=2, interpolation=cv2.INTER_CUBIC)
thresh = cv2.adaptiveThreshold(
    gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
    cv2.THRESH_BINARY, 31, 11
)

text = pytesseract.image_to_string(thresh, lang="eng")
print(text)

7) If quality is very poor

Try:

• OCR on multiple preprocessing variants and pick the best result.
• Segment the page into regions and OCR each separately.
• Use a commercial OCR engine or a cloud API.
• For handwritten text, use a handwriting-specific model rather than standard OCR.

If you want, I can give you:

• a Python script for preprocessing + OCR,
• recommendations for best OCR tools,
• or help with a specific image if you upload one.