Indexing, Object Segmentation, and Event Detection in News and Sports Videos - Multimedia Database Retrieval: Technology and Applications - page 193

where Th h ( x , y ) is the threshold parameter for pixels corresponding to the face

candidates.

7.5.1.4 Recognition Stage

The face detection stage consists of two main components: Gabor wavelets feature
extractions and AdaBoost detection algorithms. Gabor wavelets demonstrate two
desirable characteristics, spatial locality and orientation selectivity, which has
shown its effectiveness in automatic face detection and recognition. Boost algorithm
is adopted to reduce the redundancies of the high dimensional feature space and
computational cost.
The AdaBoost algorithm [ 358 ] was demonstrated to have a very low false
positive rate for face detection and can detect faces in real time. It can be trained
for different levels of computational complexity, speed and detection rates which
are suitable for specific applications. The performances of Real AdaBoost, Gentle
AdaBoost and Modest AdaBoost for face detection are compared in the current work
based on video sequences. Real AdaBoost is the generalization of a basic AdaBoost
algorithm and is treated as a fundamental boosting algorithm. Gentle AdaBoost is a
more robust and stable version of Real AdaBoost. It is shown that Gentle AdaBoost
performs slightly better than Real AdaBoost on regular data, and considerably better
on noisy data. It is also much more resistant to outliers. Modest AdaBoost is a
regularized tradeoff of AdaBoost, mostly aimed at better generalization capability
and resistance for certain specific sets of training data.

7.5.2 Experimental Result

The performance of the methods for human face detection and segmentation were
evaluated on two video datasets with different illumination conditions. The first
test video dataset was recorded under conditions of good brightness. The dataset
includes eight subjects (2 Italian, 2 Chinese, 2 Pakistani, 1 Persian, and 1 Canadian)
and comprises 520 video clips in total. The second dataset (647 clips in total)
consists of commercial films and includes videos available on the Web under
complex illumination conditions. Videos in the second dataset also contain single or
multiple faces occurring at different sizes, in different poses, and at various positions
with respect to each other.
The videos with good lighting conditions were collected for the purpose of
human emotion recognition. Each human subject showed the six fundamental
human emotional states: happiness, sadness, anger, disgust, fear and surprise.
The variations among the emotional states make the face detection task more
challenging, since the training images were essentially photographed in the neutral
state.