Detecting a face from an image is an important step to many computer vision applications, for instance extracting features from a face using a model like VGGFace2. This tutorial goes over two well known approaches: HAAR Cascade Classifier  using OpenCV and a Convolutional Neural Network, specifically MTCNN .
Disclaimer: There is a lot of debate on the ethicality of common use cases of computer vision tasks used on human faces such as emotion detection. There is an abundance of ethically questionable studies that claim to be able to tell a person's characteristics which have rightfully come under criticism. I do not wish to contribute to such work with this tutorial, and hence I do request any AI practitioners and researchers use their ethical reasoning into any application of such technology.
HAAR Cascade Classifier #
HAAR Cascading [1:1] is a technique that has been around for quite a while and yet is still very popular due to how well it works given relatively little computational power. The technique involves a sliding window over an image's pixels, very similar to a single convolutional layer in a Convolutional Neural Network. The window contains a cascade of classifier filters that have been pretrained with weights. The cascade implies that the 6000+ filters are applied one after the other, but only when previous filters have passed. This is what makes the algorithm perform so well, the sliding window only focuses on the pixels most likely to be a face.
For more details on the theory, the best source would be the OpenCV library's tutorial on this. OpenCV is a free computer vision library that includes the most well known implementation of a HAAR Cascade Classifier, which is the one we will be seeing a glimpse of today.
We'll be using the
opencv-python library (along with
PIL). You can install them if you haven't already using:
pip3 install numpy pillow opencv-python
Then in a python environment:
from pathlib import Path
from PIL import Image
import numpy as np
face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
In this bit we are initializing a Cascade Classifier. The
'haarcascade_frontalface_default.xml' is a specification for which pretrained model to use. You can find a full list of the options here.
img = cv2.imread('path/to/image.jpg')
# Convert the image to gray
gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray_image, scaleFactor=1.25, minNeighbors=3, minSize=(40,40))
The last line is where the detection happens and the parameters here are quite important. You can refer to [this stackoverflow user's answer] for more details on each parameter and recommended values for them. You may need to tinker with these values to find the best set of trade offs for your use (oh yes, this method is fast and simple, but by no means perfect).
# Check number of faces found
print('Number of faces detected:', len(faces))
# Crop out the selected faces
crops = 
for count,face in enumerate(faces):
x, y, w, h = face
left = x - padding
right = x + w + padding
top = y - padding
bottom = y + h + padding
crop = original[left:right,top:bottom]
# ensure that faces are sorted in the order in which they appear
crops.sort(key=lambda x: x['weight'])
crops = [c['img'] for c in crops]
Of course you can also draw rectangles on the faces on the original image as well.
# copy for drawing rectangles on
image_copy = np.copy(img)
for (x,y,w,h) in faces:
Why Use HAAR Cascade Classifier? #
It is very fast, and does not require a lot of computational resources. This makes it perfect for running if you don't have a GPU, or for running on an embedded devices.
Why Not Use HAAR Cascade Classifier? #
While the performance is good enough for a lot of use cases, it is also very brittle. It is not reliable for detecting a face that is not looking straight at the camera. The algorithm specializes in edge detection, and doesn't work quite when the edges are not obvious, such as in the dark or wearing sunglasses. And yes, it works particularly poorly on people not presenting their assigned gender.
MTCNN: A Better (and more expensive) Approach #
Because of the shortcomings of HAAR Cascade Classifier, several other strategies have been developed particularly using convolutional neural networks. One particular approach combines the best of both approaches- HAAR Cascade and CNN and that is MTCNN as described by Zhang et. al. [2:1]. The theory of it is quite fascinating so I do recommend reading the paper but the cliffnotes version of it is that it uses three tasks- face detection, bounding box regression and facial landmark detection to extract a face from an image.
Why Use MTCNN over HAAR Cascade? #
It is far more flexible at detecting various poses, faces with accessories and various types of faces (although it too unfortunately performs disproportionately better for light skinned face versus dark skinned face, the difference is not as stark as with HAAR Cascade). It is perfect if you intend to perform further feature extraction from the faces, for example using a model such as VGGFace2 (as you'll be needing a GPU for that anyway).
Why Not Use MTCNN? #
Unfortunately as MTCNN is much more computationally expensive to run, it needs to be run on a GPU. This makes it less ideal for usage on edge devices and embedded devices.
We will using an implementation from the Facenet library adopted to PyTorch in facenet-pytorch.
First install the package.
pip install facenet-pytorch
And run in python:
from facenet_pytorch import MTCNN
from PIL import Image
import numpy as np
img = Image.open(img_path)
mtcnn = MTCNN(margin=20, keep_all=True, post_process=False, device='cuda:0')
The parameters to the MTCNN class will represent what task you intend to perform. In my case, I want to extract all detected faces, and hence I set
keep_all=True. I am also adding a margin of 20 pixels because I want a little more of the face to be included in the crop but this may differ for you. For a more comprehensive list of options for various use cases, please refer to this notebook.
faces = mtcnn(img)
The detected faces should be of shape
[n, channels, width, height] where n is the number of faces detected. Width and height can be configured by passing a parameter to the MTCNN class.
To separate (and perhaps visualize the faces in a notebook), you can run:
fig, axes = plt.subplots(1, len(faces))
for face, ax in zip(faces, axes):
ax.imshow(face.permute(1, 2, 0).int().numpy())
In this tutorial, we have discussed two approaches for face detection, HAAR Cascade Classifier and MTCNN. We have also seen code examples of a Python implementation of each, and discussions of the advantages and disadvantages of using one over the other.
Rainer Lienhart and Jochen Maydt. An extended set of haar-like features for rapid object detection. In Image Processing. 2002. Proceedings. 2002 International Conference on, volume 1, pages I–900. IEEE, 2002. ↩︎ ↩︎
Zhang, K., Zhang, Z., Li, Z., and Qiao, Y. (2016). Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Processing Letters, 23(10):1499–1503. ↩︎ ↩︎