justinshenk/fer

Facial Expression Recognition with a deep neural network as a PyPI package

FER

Facial expression recognition.

INSTALLATION

Currently FER only supports Python 3.6 onwards. It can be installed through pip:

$ pip install fer

This implementation requires OpenCV>=3.2 and Tensorflow>=1.7.0 installed in the system, with bindings for Python3.

They can be installed through pip (if pip version >= 9.0.1):

$ pip install tensorflow>=1.7 opencv-contrib-python==3.3.0.9

or compiled directly from sources (OpenCV3, Tensorflow).

Note that a tensorflow-gpu version can be used instead if a GPU device is available on the system, which will speedup the results. It can be installed with pip:

$ pip install tensorflow-gpu\>=1.7.0

To extract videos that includes sound, ffmpeg and moviepy packages must be installed with pip:

$ pip install ffmpeg moviepy 

USAGE

The following example illustrates the ease of use of this package:

from fer import FER
import cv2

img = cv2.imread("justin.jpg")
detector = FER()
detector.detect_emotions(img)

Sample output:

[{'box': [277, 90, 48, 63], 'emotions': {'angry': 0.02, 'disgust': 0.0, 'fear': 0.05, 'happy': 0.16, 'neutral': 0.09, 'sad': 0.27, 'surprise': 0.41}]

Pretty print it with import pprint; pprint.pprint(result).

Just want the top emotion? Try:

emotion, score = detector.top_emotion(img) # 'happy', 0.99

MTCNN Facial Recognition

Faces by default are detected using OpenCV's Haar Cascade classifier. To use the more accurate MTCNN network, add the parameter:

detector = FER(mtcnn=True)

Video

For recognizing facial expressions in video, the Video class splits video into frames. It can use a local Keras model (default) or Peltarion API for the backend:

from fer import Video
from fer import FER

video_filename = "tests/woman2.mp4"
video = Video(video_filename)

# Analyze video, displaying the output
detector = FER(mtcnn=True)
raw_data = video.analyze(detector, display=True)
df = video.to_pandas(raw_data)

The detector returns a list of JSON objects. Each JSON object contains two keys: 'box' and 'emotions':

• The bounding box is formatted as [x, y, width, height] under the key 'box'.
• The emotions are formatted into a JSON object with the keys 'anger', 'disgust', 'fear', 'happy', 'sad', surprise', and 'neutral'.

Other good examples of usage can be found in the files demo.py located in the root of this repository.

To run the examples, install click for command line with pip install click and enter python demo.py [image|video|webcam] --help.

TF-SERVING

Support running with online TF Serving docker image.

To use: Run docker-compose up and initialize FER with FER(..., tfserving=True).

MODEL

FER bundles a Keras model.

The model is a convolutional neural network with weights saved to HDF5 file in the data folder relative to the module's path. It can be overriden by injecting it into the FER() constructor during instantiation with the emotion_model parameter.

LICENSE

MIT License.

CREDIT

This code includes methods and package structure copied or derived from Iván de Paz Centeno's implementation of MTCNN and Octavio Arriaga's facial expression recognition repo.

REFERENCE

FER 2013 dataset curated by Pierre Luc Carrier and Aaron Courville, described in:

"Challenges in Representation Learning: A report on three machine learning contests," by Ian J. Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, Yingbo Zhou, Chetan Ramaiah, Fangxiang Feng, Ruifan Li, Xiaojie Wang, Dimitris Athanasakis, John Shawe-Taylor, Maxim Milakov, John Park, Radu Ionescu, Marius Popescu, Cristian Grozea, James Bergstra, Jingjing Xie, Lukasz Romaszko, Bing Xu, Zhang Chuang, and Yoshua Bengio, arXiv:1307.0414.