MVA2002 IAPR Workshop on Machine Vision Applications, Dec. 11 - 13, 2002, Nara- ken New Public Hall, Nara, Japan

2-2 Analysis of Human Skin Color Images for a Large Set of Color Spaces and for Different Camera Systems

t~ean-~hristo~heTerrillon, t~rnaudPilprk, t~oshinoriNiwa and t~azuhikoYamamoto toffice of Regional Intensive Research Project (HOIP), Softopia Japan Foundation, 4-1-7 Kagano, Ogaki-City, Gifu 503-8569, Japan {temllon, pilpre, niwa)@hoip.jp t~acultyof Engineering, Gifu University, Yanagido, Gifu-City, Cnfu 501-1 193, Japan [email protected]

Abstract ultimately limits the quality of segmentation or of skin pixel detection is the degree of overlap, or discrimination, Human skin color is a powerful fundamental cue that between the skin distribution and a distribution of "non- can be used in particular, at an early stage, for the skin" pixels in a given chrominance space, which depends important applications of face and hand detection in color to some extent on the number of skin and non-skin pixels images [l] [2], and ultimately, for meaningful human- that are collected for calibration. To our knowledge, no in- computer interactions. In this paper, we analyze the depth comparative study of the efficiency of a large set of distribution of human skin for a large number of different chrominance spaces for skin color-based image chrominance spaces and for skin images recorded with segmentation has been performed, although we recently two different camera systems. By use of seven different presented preliminary results for a relatively limited criteria, we show that mainly the normalized r-g and CIE- number of spaces [3]. xy spaces, or a space constructed with a suitable linear In this paper, we analyze the distribution of human skin combination of normalized r, g and b values, are the most for a large set of 25 different color spaces (41 efficient for skin color-based image segmentation. In chrominance spaces), for facial skin images recorded with particular, they allow the use of a simple, single Gaussian two different camera systems, and in terms of seven skin chrominance model, and they yield the most robust different criteria. The color spaces considered here that skin distribution to a change of camera system. result from a linear transformation from the RGB space are the IIIZ13(Ohta's optimized color features [4]), hlh2h3 1. Introduction (Wesolkowski's color space [5]), YCblCrl (using the CIE standard illuminant C) and YCb2Cr2 (using the CIE Image segmentation based on skin color (or skin pixel standard illuminant D65) [6] [7], YES (a standard space detection) is relatively robust to changes in illumination, developed by the Xerox company), YIQ and YUV spaces. in viewpoint, in scale, to shading, partial occlusions and to The color spaces that result from a non-linear cluttered backgrounds as compared to the segmentation of transformation form a second group, that is divided into 4 gray-level images. Robustness of segmentation is sub-groups: the normalized color spaces (r-g-b [8] [9], generally achieved by separating the chrominance from CIE-xyz [8] [9] for both standard C and D65 illuminants, the luminance in the original RGB color image, and by and TSL [lo]), the perceptually plausible color spaces using only the chrominance for segmentation. This (CIE-DSH [8], HSV and HSL [I I]), the perceptually separation implies a dimensionality reduction by a uniform color spaces (CIE-L*u*v* [8], CIE-L*a*b* [8], suitable, linear or non-linear transformation from the 3-D and Famsworth's F-uv space [8] [12], for both standard C RGB color space into a 2-D chrominance space. One and D65 illuminants), and other color spaces (C1C2C3, important issue is the selection of an efficient 111213,and I1'l2'13' proposed as color invariants and used chrominance space, which motivates an analysis of human for viewpoint-invariant image retrieval and for color- skin color for different chrominance spaces: the based object recognition by Smeulders and Gevers in [I31 performance of face and hand detection depends critically [14], r.g and rg.b log-opponent space applied to color on the performance of segmentation, which in turn image indexing by Berens and Finlayson in [15], a'-b' ultimately depends on the chrominance space that is used. space applied to the extraction of skin color areas in facial In particular, for a given set of skin sample images or images by Kawato and Ohya in [16], mod-rgb space sample pixels that is collected for calibration before proposed by Tominaga in [171, PIP2 space used for the applying the segmentation, the space that is selected construction of the Fourier spectrum of the chromaticity determines the shape of the skin chrominance distribution, by Vertan et a/. in [18], and (RIG, R/B, GfB) and YUV which in turn determines the complexity of the skin spaces). The conversions from the RGB space for both chrominance model that is required in order to obtain a groups are shown in Tables 1-5. These Tables also show high quality of segmentation. The skin chrominance the boundaries of each space, as well as the dimensions distribution also depends on the various skin groups that used to calculate the discrete skin chrorninance histogram are considered (Asians, Caucasians, dark skin groups), the for each space. For all chrominance spaces considered in illumination conditions under which the color images this paper, the histogram dimensions are selected such that were recorded, and on the camera system that is used to the histogram resolution is the same for all spaces, in record the images. Finally, an important criterion that order to ensure a valid comparative study. Table 1. Linear Transformations from RGB color Table 3. Non-linear conversions from RGB color space space. into perceptually plausible color spaces.

COCOR s HISTOQRIY mE -OW-- DOUNMRI5 NoOF lNS

D=~(R+G+B) (D.S.E) E [O; 1.01

100x100

Y R=GR&0. Hd. If R=G=B=O. DEIS=B=U

G-B R=ux(RG,B) UX(RG,B)-D~(RG,B) ' .G=u.(RG,B' = { 2*ux(RG,~)~~YRG.B) R-G B=ux(RG.B 4tu.(~~.~)-m~~~3)' HSV N..ULcw&H=W6,lfB

H = { ‘me a# fer HSV ,XS,L)E (0;1.01

1.1 (RG,B) - mh(RG.B) 0,5 ux(RG.9) + mh1RG.B) ' 100x100 - L, F:: F:: [={, ,- [.I.SIR,G,B) mh(Iy+ =I. IRG.~)~' o,51 1 L= mu(RG,B) tmh(4 , 1

, rlmMmg~u,rASV+If R=G=B=l. .1( S=A=

Table 2. Non-linear conversions from RGB color space Table 4. Non-linear conversions from RGB color space into normalized color spaces. into perceptually uniform color spaces.

s MSTOQRIY COLOn SRCE MSTOQRIY @XWERsION CQUAnONS ~w~nm~NO or ans SPACE -OWEN- bOUNMRI5 NoOFlNS 116 IF!''16. x, 0.OON156 L*={ y. rgb . 100x100 903.31$] . xr 0.006856 [a].[Q] b B If R=G=B=O, r + g + b = 1 y. R+G+B #el r=+=113 m.=13L*Im'-m'.I. r*=13L*lr' .',I 200x200 (.W -.=&. .,=+ +lSY+32 CIE-w , X , 0.607 0.174 0.201 t R 8 (C Ill.) = 0.299 0.587 m,. = ', 1 [ 0.114 11 G 1 flE-L%Pv X. + I:? + 32' ".= X. + 1:: + 32. of 0.000 0.066 1.117 'B Iel, 191 +Y,Z)8rrlkCII1-KY.Z)wp.a(. d. rd-".du,. .a. a C or Do5 1n.d- x., n.982 C1E.w X 0.430574 0.341550 0.178325 370x570 Y = 0.222015 0.706655 0.071330 cm.:IY.I=, 1 ~5m.:[:]=[;E] Ds111,) '2.' ' 1.183 (nomldn) [=) 1 0.020183 0.129553 0.939180 [el.[Ql OsL*rlM) Mnut*pl.t:m*=r'=O; m nnd h me-r*#pa ,- z Ir x=~=z=o. I +J+z= 1 X+Y+Z uc x=r==ln 200x200 g' = ( g-113). r' = ( r-113) ,S.L) E [O; 1.0. IT (*W

CIE-Lt..ba t~~l , 1 ~0.008856 of 181 f(t) = { 7.78, + . t > 0,008856 1 ,g' > 0 and r'= 0 370 x 570 4 O~L*<100 100x100 (-ldn) ,g' c Oand r'= 0 &.Y,.Z,)=&,Y.,Z.) 4 -nnd 1 ,g'=Oud r50 ID.' b' rpn 2 0 , g' = 0 .od r's 0 v=zx.v=y. w== 2 or€ 10;4.01 S= [f(r12+i2)]'" F-m 100x100 4-~+~+~-~+~::+3~--21~~J+3 vr (0;0.41 181, 11 21 L 0.299R+ 0.5876 0.114B v 6Y w = + v'-u+~+w X+l5Y+3Z--2.+12,+3 Table 5. Non-linear conversions from RGB color space 2. Parameters used for the Skin Chrominance into other color spaces. Analysis COLOR SWCE Hl8TOGRAY SWCE -RSISWMUA- #XIND*RIm NOOFBINS We first define three different parameters that we use to perform a quantitative analysis of the skin chrominance. c, =*.,d..(--J-). w{G, B) 1) The Kullback-Leibler Divergence (KLD) [3] is selected to estimate the goodness of fit of the skin chrominance =1=2=, 100x100 distribution to a simple, single elliptical Gaussian. It is 1131 c3 i = 1.2.3 U R=G=B. C, = 112. defined in the discrete case as : a R=GB=~, ut c, = o IR - GI = f g,, in(Y+ f fwij ln[y I, = R-GI+~R-BI+IG-B i=l i=l j i=l 'ij (1) I, E 1% 1/21 - where Stii is considered as the "true" distribution (the 12 = 1~ BI 11 I2l3 R-GI+IR-B(+~G-B j= 1.2.3 100x100 normalized skin histogram observed in a discrete chrominance space with M x N bins) and Gii as the (131 1, = IG-4 11 + 12 + 13 = 1 R-GI+IR-BI+IG-B "estimated" or "model" distribution (the normalized ideal If R&=B a R=G=B=O, u( 1, =I13 Gaussian histogram calculated from the mean vector and (R-G) 11'= from the covariance matrix of the skin distribution in the (R-G + (R-B)' + (G-B)' 10;2B1 same discrete space). The KLD has the following IR-B properties: i) KLD r 0 and ii) if KLD=O, then S'ii = Gii. ill I'$', 12' = i=1,2,3 100x100 (R-G + (R-B)~+ (G-B)~ Hence, the lower the value of the KLD, the higher the 11 31. 11 41 ( G-B) I,' + 12' +I,' = 1 1,' = goodness of fit to the single Gaussian model. (R-G + (R-B)' + (G-B)? 2) The Normalized Histogram Intersection (HIN) is a IfR=G=B a R=G=B=O. u( 1',=1/3 measure of the overlap between two different

r-g E I--;+-[ distributions, such as the skin and non-skin distributions r- g, rg-b wb=ID(*] = I.R+I.G-?I.B SOOX#KT In the discrete case, it is defined as: B- wbE ].m;+m[ N M [I51 = R' + G'- ZB' HIN = 7 7 min ( St1, NS', ) a' E [O; 1.O] ,= ,= (2) 100x100 N M - where rvstij= NS$ / zz~s~ is the normalized non-skin histogram calculated in the same discrete chrominance nod-rgb rnsE [@LO] 100x100 space as S'ij, with M x N bins. The lower the value of the ( m-r) + (m-g) + (m-b) = 1 m-b E 10:1.01 '''1 '''1 an=c=~=o.rc m_r=m_pm_b=fi/3 HIN, the higher the degree of discrimination between the two distributions. P,Is llfi 3) The global shift S of a distribution can be calculated as: 15Oxl5O -116%p2s 216 S = f(mxl - mxz)L+ (myt - myzf If R=G=B=0, led P1=P2.0 (3) where mx=(mxl,mxzl and my=(myt,mys) are the R1Rfl~RI=GR R2=BR &=BIG R( E [@m[, 200~200 mean vectors for the skin distributions in a given If R&=B=O, ad R1=4=&=0 i = 1.2.3 chrominance space (x, y) for cameras 1 and 2. 1 .= = -0.147R-O.?R9G+0.436B 100x100 Y O.?YYR+0.587G+O.l14B -0.503~-0.72Sg+0.436 8 289. a- (&In) 3. Skin Chrominance Analysis 0.1XSr+0.473g+0.114 "-587' 114 Yw W ,,y, 0.6lSR-O.Sl5G-0.100B i-S15; 3.1 Experimental set-up Y 0.?99R+0.SX7G+0.114B 587 299. 440x440 Images of Asian and Caucasian subjects, and of 0.71 Sr-0.4151-0.100 (n~n-okln) 0.1RS~0.4731(+0.114 subjects with dark skin color, were recorded under slowly varying illumination conditions in an office environment Two separate sets of sample images used for the skin with both the SONY and the SGI camera systems. From chrominance analysis are recorded with an inexpensive the images obtained with the SONY camera, 65, 51 and SGI camera, and with a high-quality SONY DXC-9000 10 skin sample images of Asian, Caucasian, and dark camera system respectively. The seven criteria used for skin-colored subjects respectiveiy, were manually the analysis for each space are: 1) the robustness of the selected, yielding a total of 2.1 15xIOE+05, skin chrominance distribution with respect to the intrinsic 1.630xlOE+05, and 2.580xlOE+04 skin pixels for each variability of skin color (to three different skin groups), 2) respective skin group. When using the SGI camera, 11 1 its compactness, 3) its shape, 4) the degree of skin sample images of both Asian and Caucasian subjects discrimination (or the overlap) between the skin and non- were manually selected, for a total of 1.5 15x1OEM5 skin skin distributions, 5) the robustness (or "portability") of pixels. Also, 80 "non-skin" images were selected from the skin distribution to a change of camera system, 6) the various sources, mainly from the World Wide Web, relative robustness of the skin distribution to changes in producing a total of 2.6606xlOEW6 non-skin pixels. For illumination conditions, and finally, 7) the computational each image, the 24-bit RGB values are scaled between 0.0 cost of the transformation from the 24-bit RGB (NTSC) and 1.O. Achromatic pixels (including black) were space into a given chrominance space. assigned suitable values adapted to the particular color space that is considered, as shown in Tables 1-5. For each 22 space yielding negative chrominance values, a shift was a1 Asians (b1 Caucasians (c1 Dark skin color performed so that all values are positive, without any influence on the results of the chrominance analysis. Generally, the discrete, cumulative skin and non-skin histograms are calculated over an entire space, except for the CIE-L*u*v and CIE-L*a*b* spaces, whose boundaries are curved, and for the log-opponent and RIR?Rq spaces, where the range (hence the histogram dimensions) is determined empirically, by observing the skin and non-skin distributions (we used a range of [-1.0; 2.01 along both the x and y axes for the log-opponent space, and of [0.0;2.0] for the RIR~R~space). The resolution of the skin and non-skin histograms is 0.01 unit CIE-xv chrominanct. snace for all spaces, except for the CIE-L*u*v* and CIE- L*a*b* spaces, where the resolution is 1.0 unit. 3.2 Robustness to the intrinsic variability of skin color and compactness of the skin distribution H-S chrominance space (HSV color space) Figure 1. 2-D top view of the cumulative histograms in the As an example, Figure I shows the skin distribution normalized r-g (top), CIE-xy (middle) chrnminance spaces separately for each of the three classes of subjects for the and in the H-S (HSV) space (bottom) of skin sample images r-g, CIE-xy and H-S (HSV) spaces with the SONY of (a) Asian, (b) Caucasian subjects, and (c) of subjects with camera. Table 6 shows the KLD and the HIN for the three dark skin color, recorded with the SONY camera. Here, only skin groups for all spaces, when using the SONY camera. the relevant part of the histogram is shown. Table 7 shows, for each space and for both cameras, the area of the skin distnbut~onrelative to the area of the gamut of posslble colors (owing to the particular boundanes of the CIE-L*u*v*, CIE-L*a*b*, r-g-rg-b and R1R4R3spaces, the area for these spaces is not shown). Chl -Crl r- ClE-xv (C \ T-S (TSL) F-uv (C? The gamut in all spaces with rectangular boundaries, except in the CIE-DSH, HSV and HSL spaces, fills only a part of the entire space defined by the space boundaries, and its geometry depends on the space that is considered, as Figure 2 illustrates by use of the non-skin distributions 11-13 12- a'-h' m.g-m.h PI-P2 Figure 2. 2-D top view of the cumulative histograms in for a selected number of different spaces. The normalized several selected chrominance spaces of 80 non-skin sample CIE-xy and r-g spaces, as well as the a'-b', mod-gb and images collected from various sources. PI-P2 spaces yield the most robust distributions with respect to the intrinsic variability of skin color, because: 1) the KLD is consistently lower across the three skin groups than for the other spaces, and 2) the overlap between the skin groups varies typically within a relatively narrow range, between 45% and 64% for most spaces. The l1I2l3, 11'12'13' and C1-C2 spaces yield the Normalized r-r highest overlap, indicative of a higher robustness, but this advantage is offset by the large overlap between the skin and non-skin distributions in those spaces, as shown in Subsection 3.4. In almost all chrominance spaces, the distribution for the Asian subljects, who have an intermediate skin color, is the most compact, in terms of Normalized CIE-xv the relative area of the distribution, in particular in the r-g, CIE-xy, C,-C2, PI-P2,a'&' and mod-rgb spaces. 3.3 Shape of the skin distribution A few representative examples of the cumulative distribution fir the Asian + ~aicasiansubjects, obtained S (YES color snace) with both cameras, are shown in Figure 3, for the r-g, CIE-xy, E-S and CIE-u*v* chrominance spaces. Visually, the skin distribution in the normalized spaces fits well to the single Gaussian model, whereas in the un-normalized CIE-u*v* spaces, its shape is generally complex and cannot be Figure 3. 2-D top view of the cumulative histograms in the described well by a simple model. Table 8 shows the KLD r-g, CIE-xy, E-S (YES) and CIE-u*v* chrominance spaces of for all the chrominance spaces and for both cameras. skin sample images of Asian + Caucasian subjects recorded Since the KLD is consistently lowest for the normalized with the SONY camera (left column) and with the SGI CIE-xy and r-g spaces, together with the C2-C3, a'&', camera (right column). Table 6. KLD and HIN for three different skin groups Table 7. Area of the skin distributions for the three different (Asians (A), Caucasians (C) and subjects with dark skin skin groups relative to the area of the gamut of all possible color (D) ), for 41 chrominance spaces and with the SONY colon in a given color space, for both camera systems. camera.

I I A C 1 D I A/C / AID / C/D

3.5 Robustness to a change of camera system The robustness to a change of camera system can be measured as the change in the KLD, in the HIN and the global shift S of the distribution. As seen from Table 8, the change in the KLD is lowest for the r-g, CIE-xy, C2- C3, a'-b', mod-rgb and PIP2 spaces, while the overlap of the skin distributions between the two camera systems mg-mb and PIP2 spaces, the skin distribution in those (HIN skin SONYISGI) is intermediate to low for those spaces can be modeled by a single Gaussian. spaces. The highest overlap is found for the 111213, and 11'12'13' but, as for the overlap between the three different 3.4 Discrimination between skin and non-skin skin groups, for those spaces this advantage is offset by a distributions significant overlap between the skin and non-skin Table 8 shows the HIN for Asian + Caucasian subjects, distributions, and also by very large values of the KLD for for all the chrominance spaces and for both camera both cameras. The global shift S of the distribution is low systems. For both camera systems, the overlap between to lowest for the above-mentioned 6 color spaces, and is the skin and non-skin distributions is lowest for the r-g, also low for some of the spaces resulting from a linear CIE-xy, TSL, CIE-DSH, HSV, CIE-L*u*v*, CIE-L*a*b*, transformation from RGB space. C2-C3, r-g-rg.b (In-chroma), a'-b', mod-rgb, PIPz and 3.6 Robustness to changes in illumination RIR2R3 spaces. Hence, the discrimination capabilities conditions and computational cost of the color between skin pixels and non-skin pixels are highest in space transformation these spaces. The lowest discrimination is found for the 111213and l1'l2'l3' spaces. Finally, it is well known that a normalization of RGB values by (R+G+B) or of CIE-XYZ values by (X+Y+Z) reduces the most the sensitivity of the slun distribution to Table 8. KLD, HIN, and global shift of the skin distribution changes- in illumination, and a linear transformation from for Asian + Caucasian subjects, for 41 chrominance spaces, and for both the SONY and SGI cameras. the RGB space.. . or a non-linear conversion into the normalized rgb coordinates and into the CIE-xyz space is not computationally intensive compared to that into other spaces. The mod-rgb space also provides a suitable normalization 4. Conclusions In conclusion, overall, in terms of seven different criteria, the normalized r-g and CIE-xy chrominance spaces, or spaces such as the a'-b' and PIP2spaces that are constructed as a linear combination of normalized r, g and b values, offer the best tradeoff and appear consistently to he the most efficient for skin color-based image- segmentation. In particular, the use of these normalized spaces obviates the necessity to apply a complex and computationally intensive skin chrominance model in order to obtain a high quality of segmentation, as is the case with most un-normalized spaces, in which the skin distribution is complex-shaped. The C2-C3space, the mod- rgb space that also results from a suitable normalization, and to a lesser extent the RIR2R3space, are also good candidates. Owing to their particular geometry, the 111213 and 11'12'13' spaces are the least efficient for the specific problem of image segmentation based on skin color. References [I] J.-C. Tenillon, A. Pilpre, Y. Niwa, and K. Yamamoto. Robust face detection and hand posture recognition in color images for human- machine interaction. In Proc. of the Internatronal Conference on Pattern Recognrtron, Quebec City, Canada, August 2002. pp. 204-209. 121 J.-C. Terrillon, M. David, and S. Akamatsu. Automatic detection of human faces in natural scene images by use of a skin color model and of invariant moments. In Proceedrngs of the Thrrd Internatronal Conference on Automatrc Face and Gesture Recognitron, Nara, Japan, 1998. pp. 112-117. 131 J.-C. Temllon, Y. Niwa and K. Yamamoto. On the selection of an efficient chrominance space for skin color-baed image segmentation with an application to face detection. In Proceedrngs of the fh Internatronal Conference on Qualrry Control by Art~ficral I'isron, L.e Creusot, France, May 2001. Cepadues hditions. Vol. 2, pp. 409-414. 141 Y.-I. Ohta T. Kanade, and T. Sakai. Color information for region segmentation. Computer Graphics and Image Processing, 13(3):222- 241. July 1980. 151 S. Wesolkowski, M. E. Jernigan, and R. D. Dony. Comparison of color image edge detectors in multiple color spaces. In Proceedings of the Internatronal Confirenee on Image Processing, Vancouver, Canada, September 2000. 161 CIE Colorimetry. Official recommendations of the International Commission on illumination, Publication CIE No. 15.2, Second Edition, Central Bureau of the Commission Internationale de L'kclairage, 1131 T. Gevers and A W. M. Smeulders. Color-based object recognition. Vienna. Austria (1986). Panern Recognrtron. 32(3):453-464, 1999. [7] ITl-I-R Recommendation BT.709, Basic Parameter Values for the 1141 T. Gevers and A W. M. Smeulders. Content-based image retrieval HDTV Standard for the Studio and for International Programme by viewpoint-invariant color indexing. Image and lisron computing, Exchange , [formerly CClR Rec. 7091, ITU, 1211 Geneva 20, 17:475-488, 1999. Switzerland (1990). [IS] J. Berens and G. D. Finlayson. Log-opponent chromaticity coding 181 G. Wyszecki, and W. S. Styles. Color science: concepts and methods, of colour space. In Proceedrngs of the 15Ih Internatronal Conference on quantitative data and formulae, 20d edition, John Wiley, New York 1982 Panern Recognrtron, Barcelona Spain, September 2000. pp. 206-21 1. 191 R. W. G. Hunt. Measuring Colour. 20d edition, Ellis Honvood (1991). (161 S. Kawato and J. Ohya. Real-time detection of nodding and head 1101 J.-C. Tenillon, M. David, and S. Akamafsu. Detection of human shaking by directly detecting and tracking the between eyes. In faces in complex scene images by use of a skin color model and of Proceedings of the 4" International Conference on Face and Gesture invariant Fourier-Mellin moments. In Proceedings of the 14Ih Recognrtron, Grenoble. France, March 2000. pp. 40-45. International Conference on Panern Recognition, Brisbane, Australia, (171 S. Tominaga. Illuminant estimation of natural scenes from color August 1998. Volume 2, pp. 1350-1355. images. In Proceedrngs of the Internarronal Conference on Color rn [Ill J. D. Foley, A. van Dam, S. K. Feiner, and J. F. Hughes. Second Graphrcs and Image Processrng, Saint-Etienne. France, 2000. Cepadues Edition in C Computer Graphics Principles and Practice. Addison- ~ditions.pp. 35-40. Wesley, New York, 1996. [I81 C. Vertan, M. Ciuc, and N. Boujemaa. On the introduction of a [I21 H. Wu, Q. Chen, and M. Yachida. Face detection from color images chrominance spectrum and its applications. In Proceedrngs of the using a fuzzy pattern matching method. IEEE Transact~onson Panern Internatronal Conference on Color rn Graphrcs and Image Processrng. Analysis and A4achrne Intellrgence, PAMI-2 l(6): 557-563, 1999. Saint-Etienne, France, 2000. Cepadues Editions. pp. 214-218.