Experimental results
We made a lot of experiments with different face databases and with
real signal from video camera. In most cases (~85 %) face detection
system achieved quite good performance with default parameters. Main
problems occurred while detecting rotated faces, faces of people,
wearing glasses, faces with not good lighting conditions (lamp on the
floor beyond the face).
Example of successful face detection procedure is
shown on the figure below. Face region is shown with yellow circle
and eyes – with two yellow crosses.

Fig. Example of successful
face detection
Let’s say some words about the indicated
problems. Noninvariance to the turning of the head is main deficiency
of Viola-Jones detector. So we cannot
detect faces if it is rotated more than on 15º. But firstly almost
all existing algorithms of face detection have this deficiency and
secondly noninvariance to the turning is not main requirement to our
system. Faces of people wearing glasses could be successfully
recognized using alternative cascade file, trained on faces such a
people. Lighting conditions are to be not very bad, i.e. face are to
be discernible, focally and lighted.
We also tested different face detection algorithms such as
morphologically-based face parts detection and skin-color based face
detection. This algorithms work quiet well on rather good images
(i.e. high quality contrasted images of human faces without
background etc.). But when we tested these algorithms on more real
images, in many cases they couldn’t detect anything. So these
algorithms cannot be used on their own in contrast to Viola-Jones
detector.
Conclusions
Different methods of face detection were analyzed
and Viola-Jones detector was chosen as the most appropriate for out
task. Parameters of this detector were set so that quality of face
detection was good. Experiments showed some drawbacks of the
Viola-Jones detector but most of them were expected.