Deepface is a lightweight face recognition and facial attribute analysis (age, gender, emotion and race) framework for python. It is a hybrid face recognition framework wrapping state-of-the-art models: VGG-Face, Google FaceNet, OpenFace, Facebook DeepFace, DeepID, ArcFace and Dlib. The library is mainly based on Keras and TensorFlow.
The easiest way to install deepface is to download it from PyPI.
pip install deepfaceA modern face recognition pipeline consists of 4 common stages: detect, align, represent and verify. Deepface handles all these common stages in the background. You can just call its verification, find or analysis function in its interface with a single line of code.
Face Verification - Demo
Verification function under the deepface interface offers to verify face pairs as same person or different persons. You should pass face pairs as array instead of calling verify function in a for loop for the best practice. This will speed the function up dramatically and reduce the allocated memory.
from deepface import DeepFace
result = DeepFace.verify("img1.jpg", "img2.jpg")
#results = DeepFace.verify([['img1.jpg', 'img2.jpg'], ['img1.jpg', 'img3.jpg']])
print("Is verified: ", result["verified"])
Herein, face pairs could be exact image paths, numpy array or base64 encoded images.
Face recognition - Demo
Face recognition requires to apply face verification several times. Herein, deepface offers an out-of-the-box find function to handle this action. It stores the representations of your facial database and you don't have to find it again and again. In this way, you can apply face recognition data set as well. The find function returns pandas data frame if a single image path is passed, and it returns list of pandas data frames if list of image paths are passed.
from deepface import DeepFace
import pandas as pd
df = DeepFace.find(img_path = "img1.jpg", db_path = "C:/workspace/my_db")
#dfs = DeepFace.find(img_path = ["img1.jpg", "img2.jpg"], db_path = "C:/workspace/my_db")
Herein, image path argument could be exact image path, numpy array or base64 encoded image. Also, you are expected to store your facial image data base in the folder that you passed to the db_path argument with .jpg or .png extension.
Large Scale Face Recognition - Demo with Elasticsearch, Demo with Spotify Annoy
Notice that face recognition has O(n) time complexity and this becomes problematic for millions level data and limited hardware. If you have a really strong database, then you use relational databases and regular SQL. Besides, you can store facial embeddings in nosql databases. That's a better way. In this way, you can have the power of the map reduce technology. Here, you can find some implementation experiments with mongoDb, Cassandra and Hadoop.
Herein, approximate nearest neighbor (a-nn) algorithm reduces the time complexity dramatically. Spotify Annoy, Facebook Faiss and NMSLIB are amazing a-nn libraries. Besides, Elasticsearch wraps an a-nn algorithm and it offers highly scalability feature. You should run deepface within those a-nn frameworks if you have really large scale data sets.
Face recognition models - Demo
Deepface is a hybrid face recognition package. It currently wraps the state-of-the-art face recognition models: VGG-Face , Google FaceNet, OpenFace, Facebook DeepFace, DeepID, ArcFace and Dlib. The default configuration verifies faces with VGG-Face model. You can set the base model while verification as illustared below.
models = ["VGG-Face", "Facenet", "OpenFace", "DeepFace", "DeepID", "ArcFace", "Dlib"]
for model in models:
result = DeepFace.verify("img1.jpg", "img2.jpg", model_name = model)
df = DeepFace.find(img_path = "img1.jpg", db_path = "C:/workspace/my_db", model_name = model)FaceNet, VGG-Face, ArcFace and Dlib overperforms than OpenFace, DeepFace and DeepID based on experiments. Supportively, FaceNet got 99.65%; ArcFace got 99.40%; Dlib got 99.38%; VGG-Face got 98.78%; OpenFace got 93.80% accuracy scores on LFW data set whereas human beings could have just 97.53%.
Similarity
Face recognition models are regular convolutional neural networks and they are responsible to represent faces as vectors. Decision of verification is based on the distance between vectors. We can classify pairs if its distance is less than a threshold.
Distance could be found by different metrics such as Cosine Similarity, Euclidean Distance and L2 form. The default configuration finds the cosine similarity. You can alternatively set the similarity metric while verification as demostratred below.
metrics = ["cosine", "euclidean", "euclidean_l2"]
for metric in metrics:
result = DeepFace.verify("img1.jpg", "img2.jpg", distance_metric = metric)
df = DeepFace.find(img_path = "img1.jpg", db_path = "C:/workspace/my_db", distance_metric = metric)Euclidean L2 form seems to be more stable than cosine and regular Euclidean distance based on experiments.
Facial Attribute Analysis - Demo
Deepface also offers facial attribute analysis including age, gender, facial expression (including angry, fear, neutral, sad, disgust, happy and surprise) and race (including asian, white, middle eastern, indian, latino and black) predictions. Analysis function under the DeepFace interface is used to find demography of a face.
from deepface import DeepFace
obj = DeepFace.analyze(img_path = "img4.jpg", actions = ['age', 'gender', 'race', 'emotion'])
#objs = DeepFace.analyze(["img1.jpg", "img2.jpg", "img3.jpg"]) #analyzing multiple faces same time
print(obj["age"]," years old ",obj["dominant_race"]," ",obj["dominant_emotion"]," ", obj["gender"])
Age model got ± 4.65 MAE; gender model got 97.44% accuracy, 96.29% precision and 95.05% recall as mentioned in its tutorial.
Herein, image path argument could be exact image path, numpy array or base64 encoded image.
Streaming and Real Time Analysis - Demo
You can run deepface for real time videos as well.
Calling stream function under the DeepFace interface will access your webcam and apply both face recognition and facial attribute analysis. Stream function expects a database folder including face images. VGG-Face is the default face recognition model and cosine similarity is the default distance metric similar to verify function. The function starts to analyze if it can focus a face sequantially 5 frames. Then, it shows results 5 seconds.
from deepface import DeepFace
DeepFace.stream("C:/User/Sefik/Desktop/database")
Even though face recognition is based on one-shot learning, you can use multiple face pictures of a person as well. You should rearrange your directory structure as illustrated below.
user
├── database
│ ├── Alice
│ │ ├── Alice1.jpg
│ │ ├── Alice2.jpg
│ ├── Bob
│ │ ├── Bob.jpgEnsemble learning for face recognition - Demo
A face recognition task can be handled by several models and similarity metrics. Herein, deepface offers a special boosting and combination solution to improve the accuracy of a face recognition task. This provides a huge improvement on accuracy metrics. Human beings could have 97.53% score for face recognition tasks whereas this ensemble method passes the human level accuracy and gets 98.57% accuracy. On the other hand, this runs much slower than single models.
resp_obj = DeepFace.verify("img1.jpg", "img2.jpg", model_name = "Ensemble")
df = DeepFace.find(img_path = "img1.jpg", db_path = "my_db", model_name = "Ensemble")API - Demo
Deepface serves an API as well. You can clone /api/api.py and pass it to python command as an argument. This will get a rest service up. In this way, you can call deepface from an external system such as mobile app or web.
python api.py
The both face recognition and facial attribute analysis are covered in the API. You are expected to call these functions as http post methods. Service endpoints will be http://127.0.0.1:5000/verify for face recognition and http://127.0.0.1:5000/analyze for facial attribute analysis. You should pass input images as base64 encoded string in this case. Here, you can find a postman project.
Face Detectors - Demo
Face detection and alignment are early stages of a modern face recognition pipeline. OpenCV, SSD, Dlib and MTCNN methods are wrapped in deepface as a detector. You can optionally pass a custom detector to functions in deepface interface. MTCNN is the default detector if you won't pass any detector.
backends = ['opencv', 'ssd', 'dlib', 'mtcnn']
for backend in backends:
#face detection and alignment
detected_face = DeepFace.detectFace("img.jpg", detector_backend = backend)
#face verification
obj = DeepFace.verify("img1.jpg", "img2.jpg", detector_backend = backend)
#face recognition
df = DeepFace.find(img_path = "img.jpg", db_path = "my_db", detector_backend = backend)
#facial analysis
demography = DeepFace.analyze("img4.jpg", detector_backend = backend)MTCNN seems to overperform in detection and alignment stages but it is slower than SSD.
Passing pre-built face recognition models
You can build models once and pass to deepface functions as well. This speeds you up if you are going to call deepface several times.
#face recognition
models = ['VGG-Face', 'Facenet', 'OpenFace', 'DeepFace', 'DeepID', 'Dlib']
for model_name in models:
model = DeepFace.build_model(model_name)
DeepFace.verify("img1.jpg", "img2.jpg", model_name = model_name, model = model)
#facial analysis
models = {}
actions = ['Age', 'Gender', 'Emotion', 'Race']
for action in actions:
models[action.lower()] = DeepFace.build_model(action)
DeepFace.analyze("img1.jpg", models=models)Pre-trained weights of custom models will be downloaded from Google Drive source to your environment once. Download limit of my Google Drive account might be exceeded sometimes. In this case, you might have an exception like "Too many users have viewed or downloaded this file recently. Please try accessing the file again later". You can still download the pre-trained weights from Google Drive manually. You should then download the pre-trained weights to {HOME_FOLDER}/.deepface/weights folder. It won't try to download the weight file if it exists in the weights folder. You can find out your HOME_FOLDER as shown below.
from pathlib import Path
home = str(Path.home())
print("HOME_FOLDER is ",home)Pull requests are welcome. You should run the unit tests locally by running test/unit_tests.py. Please share the unit test result logs in the PR. Deepface is currently compatible with TF 1 and 2 versions. Change requests should satisfy those requirements both.
There are many ways to support a project - starring⭐️ the GitHub repos is just one.
You can also support this project on Patreon 🙏
Please cite deepface in your publications if it helps your research. Here is an example BibTeX entry:
@inproceedings{serengil2020lightface,
title={LightFace: A Hybrid Deep Face Recognition Framework},
author={Serengil, Sefik Ilkin and Ozpinar, Alper},
booktitle={2020 Innovations in Intelligent Systems and Applications Conference (ASYU)},
pages={23-27},
year={2020},
doi={10.1109/ASYU50717.2020.9259802},
organization={IEEE}
}Deepface is licensed under the MIT License - see LICENSE for more details. However, the library wraps some face recognition models: VGG-Face, Facenet, OpenFace, DeepFace, ArcFace and Dlib. Licence types will be inherited if you are going to use those models.
Deepface logo is created by Adrien Coquet and it is licensed under Creative Commons: By Attribution 3.0 License.