Indexing, Object Segmentation, and Event Detection in News and Sports Videos - Multimedia Database Retrieval: Technology and Applications - page 189

et al. [ 208 ] demonstrated that face images with the same pose, under different
illumination conditions form a convex cone. Ramamoorthi [ 209 ] and Basri and
Jacobs [ 210 ] independently used the spherical harmonic representation to explain
the low dimensionality of face images under different illumination conditions.
However, most previous face recognition and detection systems imposed strict
restrictions on the input data and worked with the assumption that the location of
the face within a frame is known. Although their works obtained good detection
results, these systems face two main limitations: the requirement for calibrated
multicameras and the restriction of usage for certain specific applications.

7.5.1 Automatic Face Detection using Optimal Adaptive

Correlation Method with Local Normalization

The key step of the current work is the incorporation of local normalization
with OAC technique to conventional classifiers for automatic face detection on
video sequences. Each frame of the input video sequence is first extracted and
regularized by local normalization. Face candidate regions are then roughly located
by OAC. The Gabor wavelets filters are applied for local feature extraction after
preprocessing. In the final step, the face region is detected through a cascade
classifier consisting of detectors with AdaBoost algorithm.

7.5.1.1 Local Normalization

Due to the fact that variant light conditions definitely cause low detection rates
and can be eliminated by illumination normalization, normalization techniques
should be well considered in an automatic face detection system so that the system
resistance can be evaluated for the most common classes of natural illumination
variations. Most methods exploited were typically characterized by relatively low
spatial frequencies. We use local normalization in this important step in order to
keep all the useful information in illumination invariant form to facilitate accurate
and robust feature extraction and detection. The local normalization composes of
the illumination compensation and the candidate selection processes.

7.5.1.2 Illumination Compensation

The illumination normalization process consists of several stages, including gamma
intensity correction (GIC), difference of Gaussian (DoG), local histogram matching
(LHM) and local normal distribution (LND). GIC corrects the overall brightness
variation of the input image g

( , ) ( , )

is the pixel location. This procedure
compensates the pixel values of an image, under unknown lighting conditions, by

x y , where x y