How theankurgoswami/face-recognition-using-facenet works
Project Type & Purpose
• This is an open‑source face‑recognition tool built on a pre‑trained FaceNet model.
• It provides two primary flows:
a. Offline image recognition (image_recognition.py)
b. Live camera recognition (live_recognition.py)
• It also includes a “training” or “data‑generation” pipeline (script/generate_data.py) that prepares new faces for recognition.
File Structure
Top‑level:
• image_recognition.py – offline inference entry point
• live_recognition.py – live‑feed inference entry point
• requirements.txt – Python dependencies
Data directory (data/):
• data/images/ – raw training images
• data/cascade/haarcascade_frontalface_default.xml – OpenCV face detector
• data/model/ – pre‑trained FaceNet weights (.h5 or .pb)
• data/arrays/embeddings.npz & vars.npz – serialized embeddings & labels
Script directory (script/):
• generate_data.py – orchestrates face detection and embedding extraction
• detect_faces.py – encapsulates OpenCV detection + cropping
• fx.py – image preprocessing utilities
Test directory (test/):
• test/ – images for offline testing
• test/predicted/ – annotated output
Read README Insights
• Uses FaceNet for embedding extraction (TensorFlow/Keras)
• Uses OpenCV Haar cascade for initial face detection
• Embeddings are stored in NumPy npz files and compared via Euclidean distance
Main Components
a. Input Sources
• File Input (data/images/, test/)
• Camera Input (live_recognition.py)
b. Face Detection Module
• script/detect_faces.py using OpenCV Haarcascade
c. Preprocessing Module
• script/fx.py – alignment / resize / normalization
d. Embedding Extraction
• FaceNet model (data/model)
• TensorFlow/Keras runtime
e. Embeddings Store
• NumPy arrays (data/arrays/embeddings.npz, vars.npz)
f. Recognition Logic
• Distance calculation (L2) against stored embeddings
• Threshold decision → identity label or Unidentified
g. Output
• Annotated images to test/predicted/
• Live video stream with overlaid labels
Relationships & Data Flow
Training/Data‑Gen Flow (generate_data.py):
Raw images → Face Detection → Preprocessing → FaceNet → Embeddings & Labels → store npz
Inference Flow (image_recognition.py / live_recognition.py):
Input (file or camera) → Face Detection → Preprocessing → FaceNet → Embedding → Distance Compare → Label assign → Annotate & Output
Architectural Patterns & Principles
• Pipeline pattern: modular stages (detect → preprocess → embed → compare)
• Clear separation: data preparation vs. inference logic
• Data persistence via flat files (.npz) for rapid lookup
• Single‑process, synchronous execution
Technologies & Frameworks to Annotate
• OpenCV (face detection, image I/O)
• TensorFlow/Keras (FaceNet model inference)
• NumPy (array storage, distance calculations)
• Python standard libraries (file system, argparse)
Essential Elements Checklist
• Components: Face Detector, Preprocessor, FaceNet, Embedding DB, Matcher
• Data Stores: Raw Images, Cascade XML, Model Weights, Embeddings.npz + vars.npz
• Interfaces: File System I/O, Camera Input, Output Directory
• Flows: Data‑Gen pipeline vs. Inference pipelines