Database Reference
In-Depth Information
For the training of the face detector, face images and non-face images are
collected from the extended Yale Database and CMU Database, which are the
publicly available face detection databases with large illumination variations. The
detector is trained to detect a face centered in a standard window with a size of
54
48 pixels.
For the local normalization method, the nonlinear histogram equalization was
applied by taking into account histogram distribution over the local window and
combining it with the global histogram distribution. Examples of the filtered results
of the original images are shown in Fig.
7.17
. By the local normalization, it can be
observed from Fig.
7.17
that the histograms of all input images are widely spread to
cover the entire gray scale. The distribution of pixels is not too far from uniform. As
a result, dark images, the histogram components of which are concentrated at the
low end of the gray scales, bright images, the histogram components of which are
biased toward the high end, and low contrast images, the histogram components of
which are narrow and centered toward the middle of the gray scale, are significantly
enhanced to give an appearance of high contrast. By applying local normalization,
an image with varying lighting conditions shows a great deal of gray level detail and
has a high dynamic range. So the system resistance to natural illumination variation
is improved.
Gabor wavelet filters with four scales and eight orientations were applied for
feature extraction. For the purpose of training the detector, a total of 15,599
subjects (8,754 positives and 6,845 negatives) were used. The detector was trained
through cascade AdaBoost classifiers. Real AdaBoost, Gentle AdaBoost and
Modest AdaBoost were compared for error checking with 200 boosting iterations.
Gentle AdaBoost returned a better face detection rate, and was selected as the
detection algorithm.
For testing, the face detection methods were applied to the two databases,
containing various practical aspects in face detection, such as changes in illumi-
nation, poses, size and various faces. Figure
7.18
shows the overall performance
of the methods using ROC curves. The detection results were obtained by setting
the window size of the local normalization to 5
×
48, and all training images are so resized to 54
×
5. The detection method
labeled as GW utilized Gabor wavelets features only, and the method labeled as
GW + LN used combined features of GW and local normalization. These methods
were applied to video data at different illumination conditions. The experimental
results demonstrated that the face detection accuracy is considerably improved
by about 10-15 % by incorporating local normalization in the critical regions of
detection rate vs. false positives. At the same time, false detection rates dropped by
approximately 15 %.
Figures
7.19
and
7.20
show the face detection results from video sequences under
good illumination conditions and bad illumination conditions. It can be observed
that all faces were detected under varying illumination conditions. The size of the
bounding box was determined using the scale of the detected face on the image.
Finally, the face detector was applied on the video sequences containing rotating
poses, varying sizes, and multiple faces. The detection rates are given in Table
7.2
.
Columns 2, 3, 4, and 5 indicate video sequences with good illumination conditions,
×
Search WWH ::
Custom Search