Image Segmentation with Eigen-Subspace Projections - Video Segmentation and Its Applications - page 36

Fig. 2.7 Simulation results of four test sequences: ( a ) Akiyo; ( b ) Carphone; ( c )News;( d ) Mother-
and-daughter sequences ( up : original images; down : extracted skin color objects)
segmentation methods. The News sequence has two newscasters that occupy two
small regions in the scene. In this sequence, the background is more complex with
one static blue screen and a large scene-changing TV screen. Actually, the face skin
regions are very small in this sequence. The MD sequence has serious shadow effect
on mother's clothes and daughter's head.
While simulation, we extract the desired samples only from the first frame of the
sequences to obtain the covariance matrix. The skin color extraction results of the
first frame from four test sequences are shown in Fig. 2.7 . The images on left column
show the original images while the right column exhibits the extracted skin color
regions. Simulation results show that those skin regions are successfully extracted
by using our algorithm.
Inspecting Fig. 2.7 , our algorithm can also identify even the small regions such
as eyes and mouth, which are with a different color sensation from the skin. In
Fig. 2.7 b, we also extract the car's roof because it has similar color as the man's
face. Applying the temporal redundancy, of course, we can remove the car's roof
by using some motion information. In Fig. 2.7 d, the extraction results of Mother's
and daughter's faces are influenced by the shadows but the main parts of skin are
revealed. Figure 2.8 shows the extraction results of clothes and hair objects. Inspect-
ing Fig. 2.8 a,b, we can extract the Akiyo's clothes and hair separately according to
different sample color. We can also extract the clothes of News and MD as shown
in Fig. 2.8 c,d. In order to verify the robustness of our algorithm, we take different
frames in Akiyo's sequence with the same transformation matrix obtained from the
first frame. From Fig. 2.9 , we find that our algorithm can mostly extract the skin
regions, in which her different expressions can be also observed.
Generally speaking, it is almost unnecessary to perform any post processing
method in our algorithm. Inspecting the simulation results, even the small features
such as eyes and mouth can be also indicated. If needed, we can also utilize the
temporal information to remove the unwanted static scenes. The signal and noise
subspaces' thresholds can be defined according to ( 2.25 ), ( 2.28 ), and ( 2.29 ). In