Privasee

Inspiration

We realised that every time we share photos online, we risk exposing sensitive information—faces, license plates, even bank-like numbers hidden in documents. Existing tools are either manual, unreliable, or cloud-based, which raises further privacy concerns. We wanted to build something simple, fast, and completely offline that protects privacy by default.

What it does

Privasee automatically detects and censors private details in images. Using OpenCV cascades, it finds faces and license plates, while OCR + spaCy NER identifies text-based personally identifiable information (PII). Users can choose between Mosaic pixelation or Gaussian blur, adjust the strength, and download a safe-to-share version instantly. We also built a Secure Photo Gallery where originals are password-protected and censored copies are easily accessible.

How we built it

Frontend/UI: Streamlit for a simple drag-and-drop interface with instant side-by-side previews.
Detection: OpenCV Haar cascades for faces and plates; Tesseract OCR + spaCy NER for names, numbers, and entities.
Censoring: Custom functions applying Mosaic or Gaussian blur to detected regions.
Security: SHA-256 hashed password system for unlocking originals in the gallery.
File handling: Separate images/ and censored/ folders with automatic processing.

Challenges we ran into

License plate cascades bundled with OpenCV are tuned for Russian plates, which reduced accuracy on other formats.
Mapping spaCy’s entity spans back to OCR word boxes required careful alignment.
Balancing speed vs. quality: Gaussian blur looked cleaner but slowed down large images, while Mosaic was faster but harsher.
Making the gallery secure while keeping the UX simple.

Accomplishments that we're proud of

Built a fully offline tool that never uploads images to external servers.
Integrated three layers of protection: faces, plates, and text-based PII.
Designed a clean, user-friendly interface where anyone can protect their photos in seconds.
Added a secure gallery feature with password-protected originals.

What we learned

Classical computer vision is still highly effective for fast, lightweight detection.
OCR + NER is powerful for text-based PII but needs thoughtful bounding-box mapping.
Security and UX go hand in hand—privacy tools must be both trustworthy and easy to use.
Even small design choices (like side-by-side comparison) make a big difference for user trust.

What's next for Privasee

Swap out Haar cascades for a YOLO or lightweight deep learning model for better license plate detection worldwide.
Add support for more PII patterns (emails, phone numbers, NRIC-like IDs).
Enable a manual brush/rectangle tool to fix false positives/negatives.
Batch processing and CLI integration for professionals who need to redact large datasets.
Package Privasee as a lightweight desktop/mobile app so anyone can protect their privacy with one click.