Beruflich Dokumente
Kultur Dokumente
111 - 914
extract as many features of the faces as possible. Some of the REFERENCES
results are given in Fig. 4, which shows that HSV color can
most successfully segment the faces, while RGB or YCbCr [I] M-H. Yang, D.J. Kriegman, N. Ahuja, “Detecting Faces in
color fails to do so depending on the image. It also shows Images: A Survey,” IEEE Trans.PAMl,vol.24.pp.34-58,2002.
that HSV color extracts more features than the others. The [2] H.A. Rowley, S. Raluja, T. Kanade, “Neural Network-Based
control of the weights seems easier for HSV color than the Face Detection,” IEEE Trans.PAMI, ~01.20, pp.23-38, 1998.
[3] R. Feraud, O.J. Bemier, I.-E. Viallet, and M. Collobert, “A
other two, since the HSV components have clear meanings. Fast and Accurate Face Detector Based on Neural Networks,”
Figure 5 shows four different cases of the optimal weights on IEEETrans. PAMI, vol. 23, pp. 42-53, 2001.
the HSV components; they are (a) iv>w,w,, (b) whw.>w, [4] H. Schneiderman and T. Kanade, “Probabilistic Modeling of
(c) iv,=iv, (d) iv,>ivhiv, and ivh>iv,. That is, they v a v Local Appearance and Spatial Relationships for Object
depending not only on the face image but also on the Recognition,.’ Proc. CVPR, pp. 45-51, 1998.
surroundings. [5] E. Osuna R. Freund, and G Girosi, “Training Support Vector
We used the three sets of the weights, W,=(l,I,l), W2= Machines: An Application to Face Detection,” Proc. CVPR,
(1,1,5) and W3=(2,2,1), to segment the 100 images in the face pp. 130-136, 1997.
detection and image retrieval system. Manually segmented [6] M.A. Turk and A.P. Pentland, “Eigenfaces for pattern
faces of the 100 images had mean values ranging from -20 to recognition,” J.Cognitive Neuroscience,vol.3.pp.71-96, 1991.
61 in the scale of 360 for H, from 0.1 1 to 0.79 for S, and from [7] B. Fr(lba, A. Emst, and C. Kilblbeck, “Real-Time Face
0.28 to 0.93 for V. So we set the color window to have the Detection,” Proc. 4Ih IASTED International Conf. Signal and
color subspace [-20,611 for H, [O.I I , 0.791 for S, and 10.28, Image Processing, pp. 497-502, 2002.
0.931 for V for the case of face detection, and [h-8, h+8] for H, [8] S. Satoh, Y. Nakamura and T. Kanade, “Name-It: Naming
[s-0.1,s+O.I] for S and [v-O.I,v+O.I] for V for the case of and Detecting Faces in News Videos,” IEEE Multimedia, pp.
image retrieval, where (h,s.v) are mean values of the input 22-35, 1999.
face image. 191 D. Wang, “Unsupervised Video Segmentation Based on
Watersheds and Temporal Tracking,” IEEE Trans. Circuits
The results are summarized in Table I for four input face
and Systems for Video Technology, vo1.8, pp.539-545, 1998.
patterns, where the first and the second patterns are generated
[IO] C. Toklu et al.. “Simultaneous Alpha Map Generation and
from No.3 image in Fig. 3 with different segmentation errors 2D Mesh Tracking for Multimedia Applications,“ Proc.
and have similar mean color values (12, 0.19, 0.69, the third KIP, vol. 1. pp. 113-116,1997
pattern generated from No.12 image has the means (39,0.79, [ I l l E Long, D. Feng, H. Peng, and W. Siu, “Extracting
O S @ , and the fourth one generated from No. 45 image has the Semantic Video Objects,” IEEE Computer Graphics and
means (16, 0.35, 0.77). It is seen from the results that more Applications, pp. 48-55.2001.
than 95 faces out of 100 are successfully segmented, that the
rate tends to increase with increasing number of the weights,
and that using the multiple weights is more crucial for image
retrieval than for face detection.
The 100 segmented images for the second pattern and
weight (2,l,l) in Table I are shown in Fig. 6. It may be seen
that some segmented faces are cut off due to the restricted
segmentation region in size around the sampled point, and
that part of the background image is included in some of the
segmented images. The latter may be insignificant for face
detection but may affect significantly for image retrieval.
Figure 7 shows an example of detecting multiple faces, where
the two segmented regions are discriminated. The face
detection for the 100 face images took about 37 sec when
using a single set of the weights on a I.2MHz Pentiumlll PC,
and the image retrieval took one to eight sec depending on the
number of the weights used and the input image.
5. CONCLUSIONS
The segmentation method using HSV color was shown to be
more accurate and easier for face images compared with RGB
or YCbCr color. Based on this method, we constructed a face
detectioniimage retrieval system. Using a few sets of the
weights on the three HSV components it was able to detect
more than 95 faces out of 100 successfully, and it could
retrieve images from the input face image in a short time.
111 - 915
Table 1. Detection rates for the 100 face images in Fig. 3 for four
input patterns, and identification results for the same data where
the color subspaces are different between the two experiments.
Fig. 6 Segmented images for the 100 face images. where the
pattern for tlSV in Fig. 4(a) was used as input face pattern.
(C) (d) Fig. 7 Example ofmulti-face detection on the 42nd frame in Fig.
Fig. 5 Four cases of optimal weights on HSV: from top-left to 3.
bottom-right, (wh, w., w,,)=(l,lO,l), (lO,l,lO), (l,l,lO),(l,l,l),
(lO.IO.l), ( l O , l , l ) , ( l , I O . l O ) for (a), (b) and (d) and ( l , 3 , l ) ,
(l0,3,l0),(l,l.3),( l , l , l ~ , ~ l O , l O , 3 ) , ( l O , 3 , 3 )(1.3.3)for(c).
,
111 - 916