From Pixels to Privacy Using OCR Microsoft Presidio to Redact Sensitive Data in ID Cards

But reality hit different.

OCR didn’t always read correctly:

• “Identity Card No.” became “iDentiry carp No.”
• NRIC like S1509088H sometimes turned into $1509088H.

At that moment, I realized: this wasn’t going to be a smooth ride. The format looked simple — one letter, seven digits, one letter. But the OCR saw $ instead of S, or swapped a 5 for an S. Suddenly, Presidio had no clue there was even an NRIC number there.

If you’re working in the .NET ecosystem and want to avoid the OCR accuracy headaches I ran into, IronOCR is worth checking out. It wraps Tesseract with built-in preprocessing filters — so things like deskewing, binarization, and noise removal don’t require separate manual steps.

using IronOcr;

var ocr = new IronTesseract();
using var input = new OcrInput();
input.LoadImage("nric_card.png");

// Built-in filters to improve accuracy on messy scans
input.Deskew();
input.DeNoise();
input.Binarize();

OcrResult result = ocr.Read(input);
Console.WriteLine(result.Text);

The preprocessing pipeline handles a lot of the image quality issues that caused my $ vs S misreads. It’s not a drop-in for Python workflows, but for .NET projects dealing with ID cards or documents, the built-in filters make the “OCR is fragile” problem a lot more manageable.

Press enter or click to view image in full size

Because of that, Presidio often failed to detect the PII. Even when it did detect, mapping back to bounding boxes sometimes failed because the OCR coordinates were off.

Press enter or click to view image in full size

Press enter or click to view image in full size

Lessons Learned from NRIC

After testing NRIC images, it became clear that the biggest challenge was the image quality itself.

• Sharp, flat, well-lit images → detection works smoothly.
• Blurry, angled, or low-light photos → OCR misreads letters and numbers.
• Regex needs to tolerate noise → small errors can break detection.

Even with advanced tools like Presidio, redacting sensitive data from images requires careful attention to how the images are captured and processed.

Key Insights

• Presidio + OCR is powerful for structured IDs. (KTP, NRIC, credit cards, passports)
• But OCR is fragile. Low-quality images make Presidio fail downstream.
• Regex flexibility is essential. Real OCR outputs are messy, so regex must handle variations.
• Image preprocessing helps. Deskewing, binarization, noise removal all improve accuracy.
• Real-world use needs fallback. Automation first, human review second (especially in compliance-heavy industries).

Practical Use Cases

• Finance: Automatically censor credit cards in receipts.
• Healthcare: Remove patient IDs from scanned forms.
• Government: Secure ID documents during verification workflows.
• Customer Support: Redact PII before tickets are shared internally.

Conclusion: Reflections on This Learning

What I learned is that privacy automation is not about being flawless — it’s about scaling protection.

Presidio shines when text is clean and structured. When combined with Tesseract OCR, it unlocks new possibilities: automatically sanitizing ID images, securing workflows, and reducing human error.

But it also teaches a deeper lesson:

• Machines are great at speed and consistency.
• Humans are still needed for judgment and edge cases.

In other words, automation doesn’t replace responsibility — it amplifies it.

And maybe that’s the true balance: letting machines handle the routine, while humans stay in charge of trust.

Appendix: Code Snippets

Text Analysis with Presidio

import re
import cv2
import pytesseract
from presidio_analyzer import AnalyzerEngine, Pattern, PatternRecognizer
from presidio_anonymizer import AnonymizerEngine
from PIL import Image

text = pytesseract.image_to_string(cv2.imread("ktp_sample.jpg"))
analyzer = AnalyzerEngine()
analyzer.registry.add_recognizer(PatternRecognizer("INDONESIAN_NIK", [Pattern("NIK", r"\b\d{16}\b", 0.9)]))
results = analyzer.analyze(text=text, entities=["INDONESIAN_NIK"], language="en")
anonymized = AnonymizerEngine().anonymize(text=text, analyzer_results=results)
print(anonymized.text)

OCR Extraction with Tesseract

import cv2, pytesseract
from presidio_analyzer import AnalyzerEngine, Pattern, PatternRecognizer
from pytesseract import Output

pytesseract.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe"
image = cv2.imread("ktp_sample.jpeg")
ocr_data = pytesseract.image_to_data(image, output_type=Output.DICT)

text = " ".join(ocr_data["text"])
analyzer = AnalyzerEngine()
nik_pattern = Pattern(name="NIK Pattern", regex=r"\b\d{16}\b", score=0.9)
analyzer.registry.add_recognizer(PatternRecognizer("NIK_ID", [nik_pattern]))

results = analyzer.analyze(text=text, entities=["NIK_ID"], language="en")
for i in range(len(ocr_data["text"])):
    if any(extracted_text[r.start:r.end] in ocr_data["text"][i] for r in results):
        cv2.rectangle(image, (ocr_data["left"][i], ocr_data["top"][i]), (ocr_data["left"][i] + ocr_data["width"][i], ocr_data["top"][i] + ocr_data["height"][i]), (0, 0, 0), -1)

cv2.imwrite("ktp_redacted.jpeg", image)

OCR + Presidio Pipeline

from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine

# OCR result from Tesseract
extracted_text = "NIK 3171234567890123, Name: Mira Setiawan"analyzer = AnalyzerEngine()
anonymizer = AnonymizerEngine()results = analyzer.analyze(text=extracted_text, language="en")
redacted = anonymizer.anonymize(text=extracted_text, analyzer_results=results)print("=== Redacted ===")
print(redacted.text)

Visual Redaction with Bounding Boxes

import cv2
import pytesseract
image = cv2.imread("ktp_sample.jpg")
data = pytesseract.image_to_data(image, output_type=pytesseract.Output.DICT)
for i, word in numerate(data["text"]):
    if word.isdigit() and len(word) == 16:  # NIK
        x, y, w, h = data["left"][i], data["top"][i], data["width"][i], data["height"][i]
        cv2.rectangle(image, (x, y), (x + w, y + h), (0, 0, 0), -1)
cv2.imwrite("ktp_redacted.jpg", image)