facereglib

This is a very simple face recognition package which adapt from deepface (basically the model and its weights is the same as deepface). I just remove the face analysis part and narrow it for only the face recognition task. You might want to visit deepface to get a complete face recognition framework.

Installation

To use facereglib, clone the facereglib repository to your working directory:

git clone https://github.com/thaitran24/facereglib.git

Make sure to install required packages in requirements.txt with:

pip install -r requirements.txt

With Docker

Update later

Features

Models

Thís package contains some face recognition models: DeepID, DeepFace, OpenFace, VGG-Face, Facenet, Facenet512, ArcFace. This package contains 2 major objects: a recognizer which recognizes faces base on builded database features of method buildDatabase() and a detector which detects faces with Google's mediapipe.

Recognition

from facereglib.facereg import recognizer

models = [
  "vggface", 
  "facenet", 
  "facenet512", 
  "openface", 
  "deepface", 
  "deepid", 
  "arcface", 
]

# initialize face recognition model
model = reconizer.Recognizer(model_name=models[1])

The model first creates the weights directory in facereglib to store weights of the model:

project
├── facereglib
│   ├── facreg
│   │   ├── ...
│   ├── utils
│   │   ├── ...
│   ├── weights
│   │   ├── model_weights.h5

You can create the weights directory then download and save the weight from GDrive. If not, the model will automatically download and save the weight for you.

The recognizer takes images as numpy array. We can verify whether 2 faces are same or not with verify();

# face verification
result = model.verify(img1, img2)

However, to perform face recognition, you need to build the database face features representation first. You should arrange the directory of the database as illustrated below:

project
├── database
│   ├── Person1
│   │   ├── person1.1.jpg
│   │   ├── person1.2.jpg
│   ├── Person2
│   │   ├── person2.1.jpg

Then the buildDatabase() method extract the features of each image and save the image path and its features in a representation file in src_path for recognition task.

database_folder = '/database/'
representation_folder = '/represent/'

# build database features
model.buildDatabase(db_path=database_folder, db_represent_path=representation_folder)

# face recognition: find the most similar face in database
df = model.recognize(img)   # return dataframe ['identity', 'distance']

# find identities: find and sort similarity in ascending order of 'distance'
df = model.find(img)

If you've already had a representation file and don't want to rebuild the database, you can initialize the model with representation file's directory:

representation_folder = '/represent/'

model = reconizer.Recognizer(model_name=models[1], db_represent_path=representation_folder)

Then you can perform find() and recognize() method normally.