Decades, skin color-based face detection algorithms have been developed by means of various color models. As for skin color characteristics, however, false positive rates would increase and detection rates
decrease when an image contains a complicated background or many
spots whose colors are similar to the skin color. In contrast, as the
appearance-based face detection method based mainly on the sliding window means scans the whole image and utilizes learning data,
it maintains high face detection rates but the computational costs are
also high. Thus, as the size of an image increases, the scanning process involves a lot of computational costs and processing time.
This study suggests a face detector that utilizes the advantages of
the skin color-based face detector and the sliding-based face detector. As for the skin color-based face detection method, a certain skin
color region is separated from the entire image by using the skin color
region classifier, and then the face is detected by using the sliding
window-based face detector, which is one of the appearance-based
face detection methods. Hence, subwindows detect a face only in skin
color regions to reduce the number of subwindows, and the size of
them is also limited by using a skin color region. As a result, at the
same detection rates of the existing sliding-based face detector, false
positive rates are decreased and the detection speed is improved.
This study consists of the following sections: Section 2 presents
explanations on the suggested skin color region classifier. Section 3
presents the sliding window-based face detector with the sliding window-based face detector briefly explained. Section 4 shows
the performance of the suggested algorithm based on experiments.
Section 5 presents the conclusion of this study.
2. Skin color region classifier algorithm
The methodology suggested in this study consists mainly of the
skin color region classifier and the sliding window-based face detector as shown in Fig. 1 below, which presents the entire flow diagram
of the suggested facial detection method. In Fig. 1, Im indicates the
color image of (NXM), Sm a binary image in application of the skin
color segmentation, Sp a binary image after the morphological image
processing, and Ipost an image after the labeling and non-face candidate rejection process. The process of combining skin color regions is
to reduce overlapping areas with the max size of a subwindow limited. Ipost is an image delivered to the sliding window face detector.
The skin color regions delivered along are face area candidates. The
max face window is the largest area among segment sections. With
the max size of a subwindow limited, performance enhancement is
expected. The following sections explain each block.
2.1. Skin color modeling segmentation
Skin color modeling requires the process of selecting used ones
among various color models. Color models include RGB, Normalized
RGB, HIS (hue, saturation, and intensity), HSL (hue, saturation, and
lightness), TSL (tint, saturation, and luminance), and YCbCr. Among
these, a YCbCr color model distinguishes luminance elements from
chrominance elements, and a linear conversion of an RGB color model
is possible, and thus is widely used for skin color models because of
simplicity and minimal operation. This study too utilizes this YCbC