In [1]:

%matplotlib inline
from pylab import *
rcParams['figure.figsize'] = 16, 6
import numpy as np

Face Detection¶

Applying Viola-Jones based face detection in OpenCV is very simple. All you need is:

Obtain a trained face detection cascade
Apply cascade in various scales

Face Cascade¶

The complete OpenCV installation comes with some very useful cascades: https://github.com/Itseez/opencv/tree/2.4.12.3/data/haarcascades Some of these cascades are for frontal face. In addition there are frontal face detection cascades that use LBP features. The cascade we will use here is: haarcascade_frontalface_default.xml

The cascade is an XML file that describes the weak classifiers learned parameters, and the HAAR features used. We will soon inspect its contents, but first lets use it.

Using the Cascade¶

In [2]:

import cv2
face_cascade = cv2.CascadeClassifier('haarcascade_frontalface_default.xml')
lena = cv2.imread('lena.png', 0) # Load as gray
rects = face_cascade.detectMultiScale(lena, 
                                      minSize=(64,64), # Smallest window
                                      maxSize=(180, 180), # Largest window
                                      scaleFactor=1.1, # Step between windoes
                                      minNeighbors=1) # how many neighbors each candidate rectangle should have
print 'found %d faces' % len(rects)

for rect in rects:
    cv2.rectangle(lena, (rect[0], rect[1]), (rect[0]+rect[2], rect[1]+rect[3]), 255)

imshow(lena); gray();

found 1 faces

In [3]:

lena_face_img = cv2.resize(lena[rect[1]:(rect[1]+rect[3]), rect[0]:(rect[0]+rect[2])], dsize=(24,24), interpolation=cv2.INTER_CUBIC)
imshow(lena_face_img);

Live face detection¶

In [4]:

video_capture = cv2.VideoCapture(0)
while True:
    state, img = video_capture.read()
    if not state:
        # 'Video is over'
        break
    rects = face_cascade.detectMultiScale(img, 
                                          minSize=(64,64), # Smallest window
                                          maxSize=(180, 180), # Largest window
                                          scaleFactor=1.5, # Step between windoes
                                          minNeighbors=2) # how many neighbors each candidate rectangle should have
    for rect in rects:
        cv2.rectangle(img, (rect[0], rect[1]), (rect[0]+rect[2], rect[1]+rect[3]), (0, 255, 0))
    
    cv2.imshow(winname='frame', mat=img)
    key = cv2.waitKey(1) # Wait for 30 milliseconds

    if key == ord('q'):
        # The key 'q' has been pressed
        break
cv2.destroyWindow('frame')
del video_capture

Analyzing the Cascade¶

The cascade can be treated as a black box which magically finds faces. Luckily the cascade as represented as a somewhat readable XML file which we can inspect.

In [5]:

import xml.etree.ElementTree
tree = xml.etree.ElementTree.ElementTree()
root = tree.parse('haarcascade_frontalface_default.xml')
features = root.find('.//features')
print 'total number of features used in cascade:', len(list(features))
print 'total stages in cascade:', len(list(root.find('.//stages')))

total number of features used in cascade: 2913
total stages in cascade: 25

lets examine the first stage

In [6]:

first_stage_weak_classifiers = list(root.find('.//stages'))[0].find('weakClassifiers')
print 'First stage weak classifiers:', len(list(first_stage_weak_classifiers))
first_stage_features = [int(x.text.strip().split(' ')[2]) for x in first_stage_weak_classifiers.findall('.//internalNodes')]

First stage weak classifiers: 9

Now let's visualize the features of the first cascade by parsing the feature part of the xml:

In [7]:

def show_feature(feature_tag):
    canvas = np.dstack([lena_face_img]*3)
    for rect_tag in feature_tag.findall('.//rects/_'):
        tokens = rect_tag.text.strip().split(' ')
        rect = [int(x) for x in tokens[:4]]
        weight = float(tokens[-1])
        cv2.rectangle(canvas, tuple(rect[:2]), (rect[0]+rect[2], rect[1]+rect[3]), color=(255,0,0) if weight > 0 else (0,0,255), thickness=1)
    return canvas

for i in first_stage_features:
    canvas = show_feature(features[i])
    subplot(3,3,i+1)
    title('Feature %d' % i)
    imshow(canvas, interpolation='nearest')
tight_layout()