📸 Photo Album Extractor

This app automatically crops and straightens photos from photo album pages (or any scanned document!).

• Initially, I tried to vibe-code an algorithmic approach, but it wasn't reliable enough for messy photo album pages lacking a white background and clear borders.
• Then, I found NerdScan, which uses the Grounding DINO model, and adopted it as a backend for this GUI. Thanks to klimentij and IDEA-Research.
• To fix photo orientation, the app uses the Deep-OAD model.

Features

• Streamlit based GUI
• ML-Powered Detection: Uses Grounding DINO for arbitrary prompt-based detection
• Two-Pass Approach: After running an initial batch detection, you can refine your prompt and settings, and retry any individual pages.
• Auto-Rotation: Automatically corrects photo orientation using Deep-OAD
• Progress Tracking: Lists and tracks completed files in a csv file
• CPU-Only: No need for specific hardware

Setup and use

• To use with Docker, run docker compose up in the root directory, and access at http://localhost:5000.
• To run on your machine, install uv, and uv run streamlit run app.py --server.port 5000
• After running your first batch, take note of any mistakes in the detection, and refine the prompt to better match your document (e.g. "an old photo.", "a polaroid.", "a single photograph.").
• If detecting multiple photos in one, try disabling overlap removal. You'll get the full set of detections, and then you can remove any duplicates manually.

Model Downloads

The application automatically downloads required models:

• Grounding DINO
• Deep-OAD

Improvement ideas:

• GPU support
• Additional post-processing (crop out borders, auto-color-balance, etc.)
• Add file tracking to GUI
• Make the Docker image smaller

With credit to:

• Klimentiy Bulygin: NerdScan
• IDEA Research: Grounding DINO
• Subhadip Maji: Deep-OAD, with ONNX weights from Chuckame