Content uploaded by Amelia Carolina Sparavigna
Author content
All content in this area was uploaded by Amelia Carolina Sparavigna on Dec 18, 2017
Content may be subject to copyright.
An image-processing analysis of skin textures
A. Sparavigna
1
and R. Marazzato
2,3
1
Physics Department, Politecnico di Torino, Torino, Italy,
2
Department of Automation and Computer Science, Politecnico di Torino, Torino, Italy and
3
Faculty of Science and Technology, Free University of Bozen, Bolzano, Italy
Background: This paper discusses an image-processing
method applied to skin texture analysis. Considering that
the characterisation of human skin texture is a task ap-
proached only recently by image processing, our goal is to
lay out the benefits of this technique for quantitative evalua-
tions of skin features and localisation of defects.
Methods: We propose a method based on a statistical
approach to image pattern recognition. The results of our
statistical calculations on the grey-tone distributions of the
images are proposed in specific diagrams, the coherence
length diagrams.
Results: Using the coherence length diagrams, we were
able to determine grain size and anisotropy of skin textures.
Maps showing the localisation of defects are also proposed.
Conclusion: According to the chosen statistical parameters
of grey-tone distribution, several procedures to defect de-
tection can be proposed. Here, we follow a comparison of
the local coherence lengths with their average values. More
sophisticated procedures, suggested by clinical experience,
canbeusedtoimprovetheimageprocessing.
Key words: 2D textures – skin texture – texture functions –
defect localisation
&2009 John Wiley & Sons A/S
Accepted for publication 5 August 2009
AQUANTITATIVE characterisation of human
skin textures is a task only recently ap-
proached by image processing. This problem is
twofold interesting. Besides the computational
modelling of skin for realistic rendering in com-
puter graphics (1), we must consider the possibi-
lity fo applying texture analysis for computer-
assisted diagnosis in dermatology (2, 3).
Image processing can analyse the texture of the
skin, that is, the appearance of its smooth surface,
and extract some data concerning its features.
These features are dependent on many factors
such as, for instance, diet and hydration, amount
of collagen and hormones and, of course, skin
cares. A gradual decline in skin is moreover
superimposed by age. As skin ages, it becomes
thinner and more easily damaged, with the ap-
pearance of wrinkles. This deterioration is also
accompanied by a darkening of skin colour be-
cause of over-absorption of the natural colouring
pigment, melanin, by the top-most cell layer in
the skin. The skin texture is also dependent on its
body location. Moreover, when using image pro-
cessing, we have to consider the fact that the
texture appearance changes with image-record-
ing parameters and with illumination and direc-
tion of view.
The task of having a quantitative evaluation of
skin features is then quite complex, as in all the
cases when image analysis must be applied to
surfaces with irregular non-periodic patterns. In
digital image processing, several methods have
been developed to classify images and define
statistical distances among them, with the aim
of deciding whether, in a set of many images,
there exist some which are close to any arbitrary
image previously encountered. Ttexture discri-
mination can be obtained by choosing a set of
attributes, the texture features, which accounts
for the spatial organisation of the image (4–7). For
skin textures, approaches based on wavelets (8),
adaptive segmentation (9) and genetic image
analysis (10) have been proposed. Bevilacqua
and Gherardi (11) and Bevilacqua et al. (12)
have faced skin topography characterisation by
processing the skin profile obtained with a capa-
citance device, to investigate the effect of ageing.
In this paper, for skin characterisation, we
propose to use the image processing procedure
previously used to investigate texture transitions
161
Skin Research and Technology 2010; 16: 161–167
Printed in Singapore All rights reserved
doi: 10.1111/j.1600-0846.2009.00413.x
r2009 John Wiley & Sons A/S
Skin Research and Technology
in nematic liquid crystals (13, 14). This processing
is suitable for images with smooth, scarcely
regular textures, such as those observed in the
microscopic investigation of certain nematic li-
quid crystal cells. The processing is based on a
coherence length analysis, described in detail in
the following section.
Method
As recorded by a CCD camera, an image is a
bidimensional array of pixels, contained in a
rectangular frame. To each pixel at an arbitrary
location P(x,y) in the image frame, we associate a
grey tone branging from 0 to 255. We then obtain
a bidimensional function b(x,y) representative of
the image intensity (brightness) distribution.
Starting from the map b(x,y), which gives the
pixel grey tone, the following calculations can be
performed.
First of all, the average intensity of the pixel
tones is determined:
Mo¼1
lxlyZlx
0Zly
0
bx;yðÞdxdyð1Þ
where l
x
,l
y
are the x- and y-rectangular range of
the frame. More generally, the k-rank statistical
moments of the image are defined in the follow-
ing way:
Mk¼1
lxlyZlx
0Zly
0
bx;yðÞMo
½
kdxdyð2Þ
With this characterisation, we are then able to
define the average values of moments for the
whole image frame. The distribution of pixel
tones is then given according to these moments.
The dispersion of pixel tones about their average
value turns out to be given by the moment with
k52, from Eq. (2).
All integrals can be calculated on the whole
image or on a window. When using the image
windowing, moments M
o
and M
k
allow to find
the position and shape of objects, because the
distribution can change for each specific window.
In images where at a first glance no particular
objects are present, we can use the same values of
moments M
o
and M
k
given by Eqs. (1) and (2), for
the whole image, supposing the image to be
characterised by only one statistical distribution.
To decide whether an image exhibits irregular
domains or localised defects, a useful procedure
is to estimate the ratio between the intensity
standard deviation and the average intensity
and fix an acceptance limit, say, for instance of
50%. Let us stress the fact that a part of the image
can be described by an intensity distribution,
which can be essentially different from the rest
of the image or from the background distribution.
In this case, it is misleading to start from the point
of view that only one distribution is enough to
describe the whole image frame and it is better to
share the image in a lattice of windows and
discuss the statistic behaviour inside each win-
dow.
In any case, among the features of a texture,
there are its homogeneity and isotropy character-
istics. In particular, to test the hypothesis of
isotropy, it is necessary to check the presence of
preferred directions in the image frame. For this
purpose, we introduced a typical length featuring
the texture size, which was very useful for the
characterisation of liquid crystal mesophases (13,
14). Instead of measuring the texture homogene-
ity by evaluating the histogram’s entropy vs. the
distance [see for instance (15)], or by calculating
the spatial organisation by means of the so-called
‘run-length statistics’ (16, 17), we compute a set of
coherence lengths defined in the following way.
Starting from an arbitrary point P(x,y) of the
figure b(x,y), along several radial directions, we
calculate the values of M
oi
(x,y) moments, that is
Mi
ox;yðÞ¼
1
lo;iZlo;i
0
bxþrsin yi;yþrcos yi
ðÞdr
ð3Þ
where index iis ranging over the radial direc-
tions, ris the radial distance from Pand y
i
is the
angle formed by the i-direction with y-axis (see
Fig. 1 for the frame of reference). Lengths l
o,i
are
the radial distances (from P) at which the values
of the moments M
oi
(x,y) on the chosen direction
saturate, within a threshold level t, the choice of
which depends on the problem under study, to
the image average moments M
o
. This is the way
to define the local ‘coherence lengths’ L
o,i
(x,y)of
the point Pin the image frame. We named these
functions as ‘coherence lengths’ because their
behaviour is similar to that of some functions
used in condensed matter physics. There, coher-
ence length is the distance over which molecular
order is maintained. The coherence length then
scales the size of ordered domains in condensed
matter.
In the calculation of functions L
o,i
(x,y), the
pixels near the image frame boundaries are not
162
Sparavigna and Marazzato
involved, because in this case it would not be
possible to estimate the coherence lengths in all
the directions (boundary effect). On the contrary,
in standard image-processing techniques (18), the
periodicity of the image, originally present or
artificially introduced by replication of the frame,
is used to overcome the boundary problem. Let
us stress the fact that moments M
oi
(x,y) are not
calculated on a window in the image frame, but
on specific directions, therefore, the method is
different from the standard statistical approach,
allowing to take into account in a natural way, the
anisotropy in the problem of texture recognition.
In our analysis, we will use 32 directions, as
shown in Fig. 1.
Actually, we can look for anomalous beha-
viours of vectors L
o,i
(x,y) as signals for the pre-
sence of a defect in the neighbourhood of point
P(x,y) in the image frame. To discuss what can be
properly considered as an anomalous behaviour
of the coherence lengths, let us introduce the
following average values of coherence lengths,
averaged over the complete frame, for each spe-
cific i-direction:
Lo;i¼1
lxlyZlx
0Zly
0
lo;ix;yðÞdxdyð4Þ
If the image frame was strictly homogeneous,
such averaged lengths should coincide with the
actual local lengths measured for all image
points. On the other hand, if the image frame
was completely inhomogeneous, the local lengths
would be well dispersed around their averages.
The same occurs when the image frame is shared
in windows, each of them characterised by a
different intensity distribution. It is acceptable
to average the coherence length over the whole
image frame when the image can be considered
as characterised by one distribution only, within a
reasonable dispersion.
Figure 2 shows the average values L
o,i
for two
images of snake skins from the Brodatz album,
which is a collection of images, actually consid-
ered as a reference collection for image-proces-
sing studies (19). These lengths represent the
distances from a generic point of the image along
the i-direction, at which the average value of the
image intensity is practically achieved, according
to the chosen threshold. The result of the calcula-
tion is proposed as a diagram showing l
o,i
in the
32 directions of Fig. 1. We can define this diagram
has the ‘coherence length diagram’. Figure 2
actually shows two diagrams obtained by fixing
two different threshold values. To obtain the
inner diagram, we use a threshold corresponding
to 50% of the ratio ffiffiffiffiffi
M
p2=Mo. The outer diagram is
obtained with 20% of the same ratio. The dia-
grams reveal preferential directions in the image
texture, that is the anisotropy of the texture.
Lengths L
o,i
are giving a visually appreciable
result with the following meaning. The diagram
of L
o,i
lengths represents the boundary of the
smallest area about a generic point in the image
frame, on which using area averaging, we obtain
the average value M
o
of the grey-tone intensity.
The diagram then represents the boundary of a
unit cell of the image, which contains the typical
features of the whole image (this is easy to see by
comparing the diagram with the images of the
snake scales). As a consequence, this unit cell has
a behaviour similar to that of the primitive unit
cell in crystal lattices (see Ashcroft and Mermin
(20), for the properties of crystal lattices). We can
also consider the cell boundary, that is the coher-
ence length diagram, as a measure of the grain
size and use it for evaluating the coarseness of the
image texture. Any object in the image frame
with a shape different from the shape of the unit
cell or any object with a different cell average
colour tone, can be considered as a defect in the
image texture.
We used snake skins to illustrate the behaviour
of coherence length diagram, because the com-
parison of this diagram with the image texture
provides immediately the meaning of it. In case
of repetitive geometric textures, as snake skin
textures usually are, the coherence length dia-
gram calculation can be implemented together
with the Fourier analysis of frequencies. As we
observed using our approach in the microscopic
investigation of liquid crystals (13, 14), it is in the
Fig. 1. Reference frame used in the calculation of average values along
the i-direction. In the figure, we show 32 directions used in our
evaluations of coherence length diagrams.
163
Image processing of skin textures
investigation of almost smooth and homogenous
images, where the Fourier analysis is scarcely
active, that the coherence lengths are truly useful.
In the next section, we will apply our image-
processing method to one of these cases, that is,
to human skin analysis. We will discuss in parti-
cular the detection of defects displayed by the
skin texture.
Results
Let us analyse with the coherence length dia-
grams some human skin textures (21, 22). A
result is shown in Fig. 3. The original map is on
the left of the figure; in the middle, we see the
coherence length curves evaluated for two differ-
ent values of the threshold, those already used to
obtain the diagrams of Fig. 2. The shape of the
two diagrams does not substantially change. The
area is changed; this is due to the fact that to be
fulfilled, a lower threshold requires a wider area.
On the right part of Fig. 3, we propose a detection
of defects; the points marked in dark grey are
considered as ‘defects’, whereas the pixels in
light grey are the normal ones. It is not a seg-
mentation procedure at the origin of these maps,
but a criterion involving the local behaviour of
coherence lengths L
o,i
(x,y).
To detect a defective region, we reasonably
assumed that a point P(x,y) in the image frame
does not belong to a defect, when the local values
of coherence lengths L
o,i
(x,y) are coincident,
within a proper threshold, with the global values
l
o,i
. The procedure is discussed with all details in
Sparavigna and Marazzato (23). We can then
prepare a map with, for instance, a red–green
layer superimposed to the original grey map. The
defects are marked in red, whereas the points
with a local behaviour coincident with the global
one are marked in green.
With a commercial software, the most common
procedure used to identify a defect is based on
Fig. 2. Coherence length diagrams (on the right) of two images of snake skin from the Brodatz album (19). The inner curve corresponds to a threshold
of 0:5ffiffiffiffiffi
M
p2=Mo, the outer to 0:2ffiffiffiffiffi
M
p2=Mo. The scale on both axes correspond to the pixel number. Note how graphs are able to indicate the texture
anisotropy. For the chosen threshold, diagrams are the boundary of the smallest area having the same features characterising the whole image frame.
164
Sparavigna and Marazzato
the thresholding of grey tone (15). This means
that the procedure is simply a check of the pixel
intensity, to see whether it is, within a fixed
tolerance, coincident or not with a specific chosen
grey tone. The processing is known as ‘image
segmentation by thresholding’ and produces
images segmented in two or more regions accord-
ing to the thresholds used. This technique is not
investigating the neighbourhood of the pixel and
then it is impossible to ascertain if it is truly
belonging to a defect or not. With the analysis
previously discussed, the detection of defects is
glocal, that is global and local, obtained by
comparing the local coherence lengths L
o,i
(x,y),
that is a local neighbourhood about the pixel
(x,y), with the global coherence length L
o,i
dia-
gram.
In the case of human skin, a defect could be a
region with a paler or darker colour or a region
with wrinkles, for instance. Fig. 3 shows an
Fig. 3. An image of the human skin texture on the left. In the middle, we see the coherence length diagrams. The inner and outer curves have thresholds
as shown in Fig. 2. On the right part of the figure, the image shows the detection of ‘defects’. Points marked in dark grey are considered as defects (see
text for explanation). Let us note that this is not a segmentation threshold procedure, as it is possible to obtain with commercial programs.
Fig. 4. The figure shows how the method determines the presence of wrinkles and spots on the human skin. The upper image refers to the forehead skin
and the lower ones to the palm skin.
165
Image processing of skin textures
almost regular texture, with a darker region. The
coherence length diagrams are indicating that the
texture is almost isotropic, and actually we have
no wrinkles. The defect map, obtained using the
previously explained procedure, places in evi-
dence the darker regions, marking them in dark
grey. In Fig. 4, other examples of defect detection
are shown. We apply the procedure to an image
of the forehead skin with wrinkles (upper part)
and to an image of the palm skin (lower figure).
Note that the method is able to map the location
of wrinkles.
Another interesting test is mapping using co-
herence length diagrams of the textures obtained
from the capacitance system, discussed in Bev-
ilacqua and Gherardi (11). In the upper part of
Fig. 5, we can see the image as it is obtained from
the capacitance system. In the lower part, we see
the same image after normalisation of the image
contrast; the capacitance image is sensitive to
different hydration and to the presence of sweat,
which give darker regions, and then a renorma-
lisation is required. The coherence length dia-
grams of the two images (without and with
renormalisation) changes because the pixel tone
distributions of the two images are different. This
difference is strongly enhanced by our procedure
of defect detection; in the right-upper part of Fig.
5, we see the dark grey areas concentrated where
the normalisation procedure must act more
strongly in changing the pixel tone distribution.
Accordingly, the right-lower image, which is the
defect map of the renormalised image, shows no
defects.
The aim of Bevilacqua and Gherardi (11) was
the development of a device to characterise the
skin topography to measure the skin profile and
the presence of wrinkles. Renormalisation of
images is then necessary for the segmentation
of skin topography, to correlate it with skin
ageing. In the case of images recorded by cam-
eras, where illumination and direction of view
are important, a similar normalisation procedure
can be quite useful too.
Fig. 5. Image (a) shows a map of the skin similar to the one that can be obtained from a capacitance system (see Bevilacqua and Gherardi (11) for
details of the capacitive method). Image (b) is the same image (a) after a normalisation of contrast. The coherence length diagrams of the two images
change because the pixel tone distributions are different. The differences between images (a) and (b) are enhanced by the procedure of defect detection.
In the upper part, we see the dark grey pixels concentrated where the renormalisation procedure must act in changing the pixel tone distribution. In the
lower image, the number of defects is strongly reduced.
166
Sparavigna and Marazzato
Conclusion
Our image analysis of the human skin texture is
based on the evaluation of the global grey-tone
distribution of the whole image frame and then
on the coherence length diagrams. These dia-
grams are also able to estimate the texture fea-
tures, such as anisotropy and coarseness.
Moreover, the coherence length diagrams can be
used to adequately describe the presence of
wrinkles, by means of a defect map. According
to the chosen statistic parameters of the grey-tone
distribution, several procedures to defect detec-
tion can be proposed. We followed a comparison
of local coherence lengths with their average
values. More sophisticated procedures, sug-
gested by clinical experience, can be easily ap-
proached.
Acknowledgement
Many thanks are due to A. Bevilacqua and A.
Gherardi for the images obtained with the capa-
citive system (11), used in Fig. 5.
References
1. Cula OG, Dana KJ, Murphy FP, Rao BK. Skin texture
modeling. Int J Comp Vision 2005; 62: 97–119.
2. Liu J, Bowyer K, Goldgof D, Sarkar S. A comparative
study of texture measures for human skin treatment.
Information, Communications and Signal Processing
ICICS ’97. Proc Int Conf 1997; 1: 170–174.
3. Pope TW, Williams WL, Wilkinson SB, Gordon MA.
Textural parameters based on the spatial gray level
dependence method applied to melanocytic nevi. Com-
put Biomed Res 1996; 29: 429–437.
4. Haralick RM. Statistical and structural approaches to
texture. Proc IEEE 1979; 67: 786–804.
5. Weszka JS, Der CR, Rosenfeld A. A comparative study of
texture measures for terrain classification. IEEE Trans
Syst Man Cybern 1976; 6: 269–285.
6. Azencott R, Wang J, Younes L. Texture classification
using windowed Fourier filters. IEEE Trans Pattern
Anal Mach Intell 1997; 19: 148–153.
7. Haralick RM, Shanmugan K, Dinstein I. Textural fea-
tures for image classification. IEEE Trans Syst Man
Cybern 1973; 3: 610–621.
8. Doi M, Tominaga S. Image analysis and synthesis of skin
color textures by wavelet transform. IEEE Southwest
Symp Image Anal Interpret 2006; 1: 193–197.
9. Phung SL, Chai D, Bouzerdoum A. Adaptive skin
segmentation in color images. Proc IEEE lnt Conf Acoust
SpeechSignal Proc 2003; 3: 173–176.
10. Nishioka M, Fukumi M, Akamatsu N, Mitsukura Y.
2004, measurement of skin texture using genetic image
analysis. Proc Int Symp Intell Signal Process Commun
Syst 2004; 1: 787–791.
11. Bevilacqua A, Gherardi A. Age-related skin analysis by
capacitance images. Proc Int Conf Pattern Recogn 2004;
2: 703–706.
12. Bevilacqua A, Gherardi A, Guerrieri R. In vivo quanti-
tative evaluation of skin aging by capacitance image
analysis. Proc IEEE Workshop Appl Comput Vision
WACV/MOTION 2005; 1: 342–347.
13. Montrucchio B, Sparavigna A, Strigazzi A. A new image
processing method for enhancing the detection sensitiv-
ity of smooth transitions in liquid crystals. Liquid Crys
1998; 24: 841–852.
14. Sparavigna A, Mello A, Montrucchio B. Pattern recogni-
tion in the microscopy of liquid crystals: description,
comparison and choice. Recent Res Devel Patterns Rec
2000; 1: 29–40.
15. Pitas I. Digital image processing algorithms. Prentice-
Hall, Englewood Cliffs, 1993.
16. Levine MD. Vision in man and machine. Mc-Graw Hill,
New York, 1985.
17. Galloway MM. Texture analysis using grey level run
lengths. Comp Graph Image Process 1975; 4: 172–179.
18. Ja
¨hne B. Digital image processing. Springer Verlag,
Berlin, 1993.
19. Randen T. The Brodatz Online Texture Archive.
Available at http://www.ux.uis.no/ tranden/
brodatz.html, Accessed at 31 August, 2009
20. Ashcroft NW, Mermin ND. Solid state physics. Saunders
College Publishing, New York, 1976.
21. Free texture maps. Available at http://freetexturemaps.
com/gallery/gallery2/v/rawtextures/people, Accessed
at 10 March, 2009
22. Free human skin texture. Available at http://
www.3dmd.net/gallery/index.php, Accessed at 10
March, 2009
23. Sparavigna A, Marazzato R. Mapping images with the
coherence length diagrams. Computer Vision and Pat-
tern Recognition (arXiv:0807.4701v1) 2008.
Address:
Amelia Carolina Sparavigna
Physics Department
Politecnico di Torino
C.so Duca degli Abruzzi 24
Torino, Italy
e-mail: amelia.sparavigna@polito.it
167
Image processing of skin textures
