🧠 Face Search with PostgreSQL & pgvector

This project demonstrates how to build a face search system using face embeddings, deep learning, and PostgreSQL vector search with the pgvector extension. Face search enables rapid and accurate similarity-based retrieval of face images from a database — useful for surveillance, access control, e-commerce, and social platforms.

📸 What is Face Embedding?

Face embedding is a fixed-length vector representation of a face. Generated via deep learning models like FaceNet or ArcFace, it allows:

• 👥 Similar faces → embeddings close in vector space
• 🧍‍♂️🧍‍♀️ Different faces → embeddings far apart

🧮 What is PostgreSQL Vector Search?

Using the pgvector extension, PostgreSQL can:

• Store high-dimensional vectors (like 512D face embeddings)
• Perform similarity search using:
  • Euclidean Distance (<->)
  • Cosine Similarity (<#>)
  • Inner Product (<=>)
• Support Approximate Nearest Neighbor (ANN) search (optional)

🛠️ System Architecture

1. 📥 Collect Face Images via DuckDuckGo
2. 📐 Extract Face Embeddings using facenet-pytorch
3. 🗃 Store Embeddings in PostgreSQL (pgvector)
4. 🔍 Search Similar Faces via SQL
5. 🖥 Deploy Web UI with Streamlit
6. 🐳 Deploy with Docker Compose

🔍 Step-by-Step Guide

Running pgvector with docker compose

cd docker
docker compose up -d

1. 📸 Collecting Face Image Data

   notebook/1. collecting_data.ipynb
   

2. 📐 Embedding Face Images

   notebook/2. embedding_faces.ipynb
   

3. 🗃️ Storing Embeddings in PostgreSQL

   notebook/3. storing_pgvector.ipynb
   

Run Streamlit UI

Streamlit run src/app.py

📚 References

facenet-pytorch: https://github.com/timesler/facenet-pytorch

pgvector: https://github.com/pgvector/pgvector

DuckDuckGo Search: https://pypi.org/project/duckduckgo-search/