You are on page 1of 20

Intelligent Data Analysis 20 (2016) S121–S139 S121

DOI 10.3233/IDA-160850
IOS Press

Hybrid adaptation for detecting skin in color


images

Michal Kawulok∗, Jolanta Kawulok, Jakub Nalepa and Bogdan Smolka


Faculty of Automatic Control, Electronics and Computer Science, Silesian University of Technology,
Gliwice, Poland

Abstract. It has been reported in many works on skin detection and segmentation from color images that skin color models
suffer from low specificity and high variance of the skin color, and this problem can be addressed by conforming the skin model
to a presented scene. Here, we introduce a new hybrid adaptation system which combines two strategies, namely (i) adaptation
from a detected facial region and (ii) a self-adaptive scheme that creates a local model based on the response obtained using
the global one. As a result of this hybrid adaptation, we obtain a local skin color model and we use it to extract seeds for the
geodesic distance transform that determines the boundaries of skin regions. The results of our extensive experimental study
confirm that the proposed algorithm outperforms several state-of-the-art methods, as well as our earlier adaptive skin detectors.

Keywords: Skin detection, skin segmentation, adaptive color model, spatial analysis

1. Introduction

Skin detection from digital images is an important and challenging problem in computer vision and
intelligent data analysis, which has been actively explored over the years. A general goal here is to take
a binary decision concerning the skin presence for a pixel, a region of interest, an image or a video se-
quence. In case of the positive detection outcome for an image, the pixel-level detection can be improved
using image segmentation techniques that benefit from the spatial properties of skin pixels. Potential ap-
plications of skin detection and segmentation include hand detection and tracking for gesture recognition
and human-computer interaction [51,52], objectionable content filtering for the sake of blocking nude
images and videos [49], feature extraction for content-based image retrieval [48], image coding using
regions of interest [17], skin lesions segmentation [13,29], and potentially many more. Importantly, the
algorithms developed for segmenting skin regions may be applied to other computer vision tasks that
involve region-based object detection in digital images, including medical imaging or satellite data anal-
ysis.
State-of-the-art approaches to detecting skin in digital images are based on the premise that the skin
color can be modeled in various color spaces [62], which allows for discriminating between skin and
non-skin pixels based on their position in the color space [32]. In general, these methods make it possible
to assign the skin-presence probability to every pixel, independently from its context, which transforms
an input color image into a skin-presence probability map (SPPM).


Corresponding author: Michal Kawulok, Faculty of Automatic Control, Electronics and Computer Science, Silesian Uni-
versity of Technology, Gliwice, Poland. E-mail: michal.kawulok@polsl.pl.

1088-467X/16/$35.00 
c 2016 – IOS Press and the authors. All rights reserved
S122 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

Accuracy of such pixel-wise classification schemes is limited by high variance of the skin color and its
low specificity. Skin and non-skin pixel colors overlap each other regardless of a color space used, which
makes it impossible to perfectly discriminate between them exclusively based on this feature. However,
this overlap may be substantially decreased, if the skin color model is adapted to a particular image
or a video sequence. The adaptation may be based on a provided skin sample, acquired automatically
using face [33,45,63,67] or hand [9] detectors. In general, from a detected skin sample, a local skin
model is learned, combined with the global model and applied to the image. Furthermore, while it is the
chrominance, which is the most discriminative skin-presence feature, it is beneficial to extract textural
and spatial features from the input image – this increases the discriminative power of skin detectors.
In our earlier work [39], we proposed to apply the geodesic distance transform (GDT) to segment
skin regions, and the seeds for the transform are selected based on a high skin-presence probability
threshold. Later, we used the local skin color model learned from a detected facial region to extract the
adaptive seeds for the GDT [33]. The method greatly improves the skin detection scores, however its
main disadvantage is that the algorithm may get adapted to some non-skin objects, if they appear within
the facial region.
In [42], we have introduced a self-adaptive scheme to detect the seeds for the GDT. Contrary to our
face-based method [33], we train a local skin color model from the skin samples acquired by analysing
an SPPM obtained using a global model, hence we do not require any explicit skin sample be provided.
The local model is subsequently used to determine the seeds for the GDT, which is applied to segment
the skin regions as in [33]. We demonstrated [42] that although the obtained results are worse than
relying on skin samples acquired using a face detector, the new algorithm outperforms other state-of-
the-art methods, while it can be applied to the images, in which the faces are not present or have not
been detected automatically.

1.1. Contribution

In the work reported here, we propose a hybrid adaptation that combines two aforementioned ap-
proaches, namely (i) the face-based adaptation [33] and (ii) the techniques used to extract the self-
adaptive seeds [42]. If a face is detected in the image, then we analyze the SPPM within the detected
facial region to select the skin samples for the adaptation. This makes it possible to reject the non-skin
pixels, should they appear within the facial region, and in this way we effectively address the weakest
point of the face-based adaptation. Subsequently, we append these skin samples to those extracted from
the entire SPPM. From them, a local skin color model is build and applied to extract the seeds for the
GDT. Afterwards, we follow the workflow developed and described in [42]. In the paper, we report the
results of our extensive experimental study that confirm the theoretical advantages of the hybrid adap-
tation. The obtained results clearly indicate that the new method outperforms other state-of-the-art skin
detectors, as well as our earlier adaptation schemes.

1.2. Paper structure

The paper is structured as follows. In Section 2, we report the state of the art concerning skin detection
and segmentation. The baseline techniques exploited in our study are outlined in Section 3. The proposed
algorithm for skin detection is described in Section 4, the experimental study is reported in Section 5,
and the paper is concluded with Section 6.
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S123

2. Related literature

There are a plethora of skin color models [32,44], which in general can be categorized into (i) those
that are defined using a set of rules and thresholds [14,15,47], and (ii) those generated using machine
learning methods. The former are characterized with definitely lower detection accuracy than the latter,
however taking into account their low time and memory requirements, they can be efficiently deployed
in certain controlled environments. The methods that engage machine learning, require a training set
containing pixels assigned to the skin and non-skin classes. From such a labeled set, the decision rules are
learned, so as to classify every pixel of any presented image. The Bayesian classifier is highly effective
here, provided that the training set is sufficiently large [31], while for smaller sets, Gaussian mixture
models (GMMs) are often used [11,24,55] due to their higher generalization capacities. Alternatively,
the number of histogram bins per channel may be reduced for the Bayesian classifier. This has been
extensively studied in our recent survey [44]. Furthermore, support vector machines [26,43,53], artificial
neural networks [8,58], and random forests [46] are among the machine learners successfully applied to
skin color modeling.
Skin detection techniques have also been developed to segment skin lesions for the sake of der-
moscopy. This is an active research topic of medical imaging, and many methods have been developed
over time [59]. Segmentation of skin lesions may be performed taking advantage of the skin homogene-
ity in the domain of color, luminance or texture. They include statistical region merging [13], dynamic
programming [1], and wavelet-based texture analysis [12]. The segmentation phase is followed by shape
analysis to investigate the lesion type [29]. In general, these methods are specialized to deal with the
dermoscopy images – it is therefore assumed that a given image presents human skin with some lesions
that should be segmented from the background.
Skin and non-skin pixels overlap each other in color spaces, which imposes severe limitation on the
pixel-wise classification schemes based on skin color models. If the model defines the skin color so as to
omit the overlapping values, then many skin pixels are classified as background, decreasing the recall.
On the other hand, if the model considers these overlapping values as skin, then the number of false
positives is increased. It is worth noting that the overlap may be reduced, if the skin model is adapted
to individuals who appear in a presented scene [69]. Given constant lighting conditions and a limited
number of individuals in the image, skin color specificity is definitely higher than in the general case,
and overall the skin regions can be better separated from the background.
The existing adaptation methods either require a skin sample, from which the local skin model is
learned on the fly, or they use some features extracted from an input image to fit the model. In the latter
case, some methods [49,66] exploit artificial neural networks (ANNs) to select the model or its param-
eters that are predicted to deliver the best result for a presented image. In [49], chroma histograms of
skin pixels for all of the images in a training set were merged and grouped into five major chroma clus-
ters, corresponding to different lighting conditions. Subsequently, a multilayer perceptron was trained to
decide which chroma cluster should be applied to a particular image.
The most common approach to adapting color models is based on a skin sample, which may be
acquired based on tracking skin-like objects in video sequences [19], or relying on face [33] or hand [9]
detection. From such a skin sample, a local model is learned [4,23,35] with the methods that do not
require time-consuming training. It has been reported that the local model allows detecting the skin with
high precision, but the recall is often low. To address this problem, it is beneficial to combine the local
model with the global one, and the final skin-presence probability is computed as a weighted mean of
the probability values resulting from the both models. Apart from that, a global skin color locus may be
S124 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

imposed to restrict the adaptation [23,61]. In our earlier work [33], we proposed to combine the local
and global models by incorporating them into a spatial analysis framework – the local model is used
to determine the seeds for the GDT. A similar approach was proposed in [45], where the local model
is learned from the facial region. The model is used to obtain the foreground weights for the graph-cut
image segmentation, and the background weights are obtained using the global skin color model. A skin
sample may also be exploited to set the values of acceptance thresholds [9,50]. Recently, it has been
proposed [67] to use skin samples provided by a face detector for adapting the acceptance thresholds
in a single-dimensional error signal space (ESS) [14]. ESS is obtained from RGB, and skin color is
modeled using a single Gaussian.
So as to increase the discriminative power of skin color models, they may be supported with some
textural features, for example extracted using the gray-level co-occurrence matrix [65] or wavelet trans-
form [12,57]. Simple textural features were used to improve a number of skin color modeling techniques,
including ANNs [64], non-parametric density estimation of skin and non-skin classes [68], GMMs [54],
and many more [2,18,21,22]. In our earlier work [38], we found it beneficial to extract the textural
features from SPPMs rather than from the color images.
A natural way to improve the skin detection accuracy at a pixel level is to take advantage of the
fact that skin pixels are usually grouped in the spatial domain, and the skin regions are often of an
elliptical shape [48]. General-purpose image segmentation methods can be applied here, for example
those based on combined Markov random fields [6], finite elements [10,28,60], probabilistic bottom-
up aggregation [3]. Among the techniques adopted for this purpose, it is worth to mention a threshold
hysteresis [4,5], conditional random fields [16], and cellular automata [2]. Apart from that, a number of
specific algorithms devoted to segmenting skin regions have been developed. Following the controlled
diffusion method [20], at first the diffusion seeds are formed by those pixels, whose skin-presence prob-
ability exceeds the seed threshold (TαP ). Then, the neighboring pixels are iteratively adjoined to the skin
region, if they meet the diffusion process criteria, provided that their skin-presence probability is larger
than the lower-bound propagation threshold (TβP ). In our earlier works, we proposed to use the GDT to
segment skin regions [36,37,39], and we found it effective to compute the local costs for the GDT in a
combined domain of hue, luminance and skin-presence probability [39].
It has been proven beneficial to combine the aforementioned improvement strategies. In [30], SPPMs
obtained using a color model are refined relying on textural features extracted using the Gabor wavelets,
and finally the watershed segmentation is applied to exploit the spatial properties of skin regions. Com-
bining textural features with spatial analysis was also the key contribution of our recent work [40]. We
introduced the discriminative skin-presence feature (DSPF) space, which is used to compute the local
costs for the GDT, instead of using the SPPM as in [39].
Recently, we proposed a self-adaptive algorithm for segmenting skin regions [42], which is based
on the idea of using a local skin color model to extract the seeds for the GDT [33]. The advancement
consists in using the SPPM extracted with a global skin color model to acquire the skin samples that are
utilized to train the local model. This makes it possible to adapt the method without relying on a face
detector. Furthermore, we proposed to compute the local costs for the GDT in the DSPF space [41].

3. Baseline methods

In this section, we present our earlier techniques for spatial analysis and model adaptation that are
utilized in this study, namely: (i) spatial analysis using the geodesic distance transform in a combined
domain (DTCD) [39], (ii) GMM-based adaptive seeds [33], and (iii) the DSPF space for computing the
local costs for the GDT [40].
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S125

3.1. Geodesic distance transform in a combined domain

Our DTCD algorithm [39] requires the skin seeds be extracted at first. From these seeds, the “skinness”
(i.e., the likeliness to the skin) is propagated using the GDT. Proper extraction of the seed pixels is
a critical operation in this algorithm, because if no seed is positioned inside a skin region, then the
entire region will obtain very low “skinness” values and will be classified as background. Therefore,
not only should the seeds contain no false positive pixels, but also every ground-truth skin blob (i.e., a
region composed of the real skin pixels) should include at least one detected seed inside. Originally, we
adopted an approach applied in many spatial analysis methods [4,5,20] to build the seeds from the pixels
with high skin-presence probability that exceeds a threshold TαP .
In order to propagate the “skinness” from the seeds, the shortest routes from the seed to every pixel
are determined at first. This is achieved by minimizing total path costs from the set of seed pixels to
each non-seed pixel in the image. The total path cost for a pixel x is defined as [39]:


l−1
Γ(x) = γ (pi → pi+1 ) , (1)
i=0

where γ is a local propagation cost between two neighboring pixels, p0 is a pixel that lies at the seed
boundary, pl = x, and l is the total path length. The minimization is performed using the Dijkstra’s algo-
rithm [27]. In addition, the threshold TβP = 0.3 is used as proposed in [20], which prevents propagating
to the regions of very low skin-presence probability. Furthermore, mainly to decrease the execution time,
the propagation is terminated, if the total path cost exceeds a certain boundary value TΓ .
The route optimization outcome heavily depends on how the local costs γ are computed. For skin
segmentation, we construct the local cost using two major components, namely the difference in the
propagation domain γΔ and the destination-probability cost γp . The local cost from a pixel x to y ,
i.e., γ (x → y) is obtained as:

γ (x → y) = γΔ (x, y) · [1 + γp (x → y)] , (2)

where


⎨−1 for P(y) > T0P
γp (x → y) = 1 − P(y) for TβP < P(y)  T0P . (3)


∞ for P(y)  TβP

P(y) is the skin-presence probability of the pixel y and T0P is the costless propagation threshold (if the
skin-presence probability at the pixel y exceeds T0P , then the total path cost does not increase when
moving from the pixel x to y ). The difference cost γΔ was originally defined using hue and luminance
values [39]:

γΔ (x, y) = αd · (|Y (x) − Y (y)| + |H(x) − H(y)|) , (4)



where αd ∈ {1, 2} is the penalty for propagation in the diagonal direction, Y (·) is the pixel luminance
and H(·) is the hue in the HSV color model, both scaled to the range from 0 to 255.
The total path cost obtained after the optimization is inversely proportional to the “skinness”, hence
the final SPPM is obtained by scaling the costs from 0 (for the maximal cost) to 1 (for a zero cost, i.e.,
the seed pixels).
S126 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

3.2. Face-based adaptive seeds

In [33], we proposed to create the seeds using a local model learned from the pixels positioned within
a detected facial region. Face and eye detection is proceeded relying on the geometrical features ex-
tracted from the luminance channel of the input image [34]. Based on the detected position of the eyes,
we determine a facial region, which is supposed to contain exclusively skin pixels. We model their dis-
tribution for every image with a single multivariate Gaussian. Afterwards, the skin-presence probability
is computed for every pixel in the input image based on its color c in RGB:
 
PS (c) = exp −0.5(c − c)T Σ−1 (c − c) / (2π)3 |Σ|, (5)

where c is the mean color value and Σ is the color covariance matrix of the pixels from the training
set. This transforms an input image into the SPPM, termed local SPPM, as it is obtained using a locally
learned skin color model. Such a local SPPM classifies the skin with high precision, but low recall, and
therefore it is suitable to determine the seeds for the GDT.

3.3. Discriminative skin-presence features domain

The DTCD method [39] uses the skin-presence probability obtained from a global skin color model
to compute the destination-probability cost γp (Eq. (3)). However, later we proposed to refine the skin-
presence probability relying on the textural features [40], and to use the refined probability for computing
the local cost γp .
The textural features are incorporated into the DSPF space, exploited to refine the skin-presence prob-
ability. In order to obtain the DSPF space, the basic image features are first extracted from the SPPM.
They consist of the following features: (i) the median and (ii) minimal values, (iii) standard deviation,
and (iv) the difference between the maximum and minimum, computed in three kernels: 5 × 5, 9 × 9, and
13 × 13 pixels. In addition, the raw skin-presence probability value is appended to this feature vector, as
it is the principal source of the discriminating information between skin and non-skin pixels. Overall, ev-
ery pixel x is transformed into a 13-dimensional basic feature vector. Using linear discriminant analysis,
we define the DSPF space that reduces the dimensionality of the basic image features and increases their
discriminative power. A similar approach has been recently applied to combine more advanced textural
descriptors [7], but here we found such simple statistical features more suitable.
Subsequently, a pixel of the maximum skin-presence probability is found in the SPPM eroded using
a large (15 × 15) kernel – it should be larger than the kernels used for extracting basic image features,
so as to include only skin pixels. This pixel is termed the reference pixel r , and the distance between
r and every pixel in the image is computed in the DSPF space. This operation converts the input skin-
presence probability map into the DSPF skin map, which is normalized and used for computing the
destination-probability cost γp .

4. Skin segmentation from hybrid-adaptive seeds

In this section, we describe our contribution that consists in combining two approaches towards ex-
tracting adaptive skin seeds for the GDT. We have demonstrated in [42] that the best segmentation results
are obtained using the face-based adaptive seeds [33], provided that the eyes are correctly localized. That
method (termed fa-DTCD), as well as other existing face-based adaptive algorithms mentioned earlier
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S127

in Section 2, is very vulnerable to false positive face detection, imprecise eye localization and non-skin
objects appearing within the facial regions. This may make the local model adapted to the background
pixels, which severely affects the achieved detection scores. So as to address this limitation, we intro-
duced a self-adaptive algorithm (sa-DTCD) [42] which analyzes the SPPM obtained using a global skin
color model (termed global SPPM) to (i) decide whether an image presents human skin at all, and (ii) in
the case of the positive answer, to extract some skin samples for training a local skin color model (same
as the facial regions in fa-DTCD). In this way, we obtain the seeds for the spacial analysis using DTCD.
Later, in [41], we reported the results obtained using the DSPF space that takes into account the textural
features (this approach is termed sa-DSPF).

4.1. Hybrid adaptation outline

In this paper, we introduce a hybrid adaptation system, which allows us to benefit from the additional
information provided by face detectors, but at the same time it provides protection against incorrect
adaptation. Contrary to fa-DTCD [33], we do not assume that the facial region is composed exclusively
of skin pixels. Instead, we analyze the global SPPM around the detected facial region to extract skin
samples using the self-adaptive techniques from the sa-DTCD algorithm [42]. In addition, we extract the
skin samples from the entire image, however the acceptance conditions are much stricter here than for
the facial region. Basically, we assume that the majority of the pixels within the facial region present the
skin, while only some pixels positioned outside the facial regions may belong to the skin. This approach
is superior to fa-DTCD because of two reasons: (i) it allows rejecting some non-skin pixels that may
appear within a facial region (it also makes the adaptation less sensitive to imprecise face detection),
(ii) the skin samples may be acquired not only from the facial region, which prevents the adaptation
from being too specific and reduces the false negatives. The advantage compared with sa-DTCD and
sa-DSPF consists in improving the adaptation capacities by utilizing the data concerning face location,
if they are available.
A flowchart that presents the whole workflow is presented in Fig. 1 and the process is illustrated using
two examples in Fig. 2. The algorithm requires a global skin color model to transform an input color
image into a global SPPM (Fig. 2(b)). From this map, we extract the initial skin seeds (black pixels
inside the gray regions in Fig. 2(c)), optionally using the information on facial regions location. These
initial seeds are slightly expanded to include more skin pixels (gray regions in Fig. 2(c)), from which a
local skin color model is trained. This model is applied to the image, and thus the local SPPM is obtained
(Fig. 2(e)), from which the adapted skin seeds are extracted. Finally, the “skinness” is propagated either
using the DTCD or DSPF algorithm to determine the boundaries of detected skin regions (Fig. 2(f),
where red indicates false positives and blue – false negatives). In Fig. 2(g), we present the segmentation
result obtained directly from the global SPPM. For the image in the upper row, the adaptation reduced
false positives (the background objects have skin-like color), while in the bottom row, false negatives
were decreased (the skin does not have a typical color here due to a flash light).

4.2. Initial skin seeds extraction

Extracting the initial skin seeds is a critical step in our algorithm. We intend to eliminate non-skin
pixels in the initial seeds, while at the same time we do not aim at finding an initial seed in every skin
region. The general goal here is to obtain a representative skin sample at the smallest possible false
positive rate.
S128 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

Fig. 1. Flowchart of the proposed skin segmentation process.

(a) (b) (c) (d) (e) (f) (g)

Fig. 2. Subsequent steps of our method: (a) input color image, (b) global SPPM, (c) ISSs before (black) and after (gray) the
DT-based expansion, (d) seeds after adaptation, (e) final SPPM, and segmentation result from (f) the adapted and (g) global
SPPM.

The procedure for extracting the initial skin seeds is presented in Algorithm 1, and it may be executed
for the entire image or restricted to a facial region MF . First, we compute the integrated histogram
H (line 1) of the global SPPM (PS ) to find the value of a dynamic threshold tseed that selects Rseed
pixels of skin-presence probability exceeding tseed (line 2). If the image contains skin pixels at all,
then the dynamic threshold is used to binarize the SPPM and extract the initial seeds (line 4). We have
observed that when the entire SPPM is considered, then the false positive pixels are scattered in the
binarized image, while the true positive pixels are organized in spatially consistent groups. Following
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S129

Algorithm 1 Initial seeds extraction.


Require: PS , (MF )  Global SPPM and (alternatively) a face mask
1: H ← I NTEGRATED H ISTOGRAM(PS );
2: tseed ← F IND T HRESHOLD (H, Rseed );
3: if I S S KIN P RESENT (PS ) then  Skin detection at the image level
4: S 0 ← B INARIZE(PS , tseed );  PS is binarized with tseed to extract the seeds
5: if MF = then  No facial region provided
6: B ← F IND B LOBS(S 0 );
7: B ← F ILTER B LOBS(B );  10% of the largest blobs are selected
8: S 0 ← R ENDER(B );
9: S 0 ← E RODE(S 0 , 11);  Isolated small blobs are removed
10: end if
11: S 0 ← M ORPHOLOGICAL S KELETON(S 0 );  S 0 contains the initial seeds
12: end if

(a) (b) (c) (d) (e) (f)

Fig. 3. Examples of facial regions used in [33] (top row) and in the proposed method (bottom row), along with the selected
initial skin seeds (red pixels in the bottom row).

this observation, we use only 10% of the largest blobs (line 7). These blobs are additionally eroded
(line 9) to eliminate the blobs having small area. Finally, the seeds are subject to the morphological
skeletonization (line 11), which further reduces the false positives.
The ratio of the pixels used to determine the value of the dynamic threshold is Rseed = 50% for
the facial region, Rseed = 1% for the entire image, if some facial regions have been detected, and
Rseed = 5% otherwise. Two examples of the initial skin seeds extracted from the entire SPPM have
been presented in Fig. 2(c). In Fig. 3, we show the facial regions and the initial seeds extracted when the
faces are detected in the image. In the top row, we demonstrate the skin regions that are entirely used to
build a local model in fa-DTCD. In the bottom row, the red pixels indicate the extracted initial seeds. For
some images (b, c, f) it may be observed that the seeds have also been found outside the facial region
(after processing the entire SPPM with the ratio Rseed = 1%). It is worth noting that for virtually all of
the cases presented in Fig. 3, the facial regions contain some background objects, and the initial seeds
are correctly picked exclusively from the skin areas. The skin regions used in fa-DTCD (top row) are of
a smaller size, but still some of them (c–e) include the non-skin pixels that jeopardize the adaptation.
An important decision in extracting the initial seeds is whether an image presents the skin at all. Here,
S130 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

we adopted the same strategy as in sa-DSPF – an image is regarded as presenting some skin regions if
(i) tseed  0.12 and (ii) a value of the reference pixel r (determined as explained earlier in Section 3.3)
in the dilated SPPM (with a 15 × 15 kernel) is not below 0.24. These parameters controlling the image-
level skin detection have been tuned on the experimental basis, which is thoroughly justified in [41].
Naturally, if the initial seeds are extracted from a facial region, then we assume that the skin is present
there.
4.3. Seeds expansion and adaptation
The initial skin seeds indicate skin regions with very high precision (i.e., they contain very few false
positives), but they are not present in every ground-truth skin blob (see Fig. 2(c)). The former property
makes the seeds appropriate to initiate the GDT in order to determine the boundaries of those skin re-
gions, in which the seeds appear. However, the second property means that the “skinness” cannot be
propagated in the spatial domain to non-covered skin blobs, hence the color space must be used for
transferring the “skinness”. This transfer is achieved by modeling the skin distribution using a single
multivariate Gaussian, and applying the model to the entire image. After this operation, the seeds are ex-
pected to appear in every skin blob, and they are used for the final spatial analysis using the DTCD algo-
rithm [39]. Here, we consider two variants of our method, depending on how the destination-probability
cost γp is computed: (i) using the skin-presence probability values from the global SPPM (termed ha-
DTCD) and (ii) using the skin-presence probability refined within the DSPF space (ha-DSPF).
Before learning the local model, the initial skin seeds are enlarged to include more skin pixels and
increase the representativeness of the local training set. The seeds are expanded using the DTCD algo-
rithm. To minimize the number of adjoined non-skin pixels during the expansion, we compute the local
cost in a 4-neighborhood and we ignore the pixels, whose total cost is above TΓ = 3γΔ , where γΔ is
the average local cost in the image. Following our earlier study [41], we use three different metrics for
computing the difference cost component γΔ :
(HL)
γΔ (x, y) = αd · max (|Y (x) − Y (y)| , |H(x) − H(y)|) , (6)
(RGB)
γΔ (x, y) = αd · (|R(x) − R(y)| + |G(x) − G(y)| + |B(x) − B(y)|) , (7)
(SP )
γΔ (x, y) = αd · |PS (x) − PS (y)| . (8)
(SP)
In the case of the ha-DTCD variant, we use γΔ
to enlarge the initial seeds prior to the adaptation,
(RGB) (HL)
and γΔ for the final “skinness” propagation. For ha-DSPF, the γΔ cost is applied in both cases.
After propagation, the distance map is normalized, as stated in Section 3, to obtain the final SPPM.

5. Experimental validation
In our experimental study, we used the ECU benchmark set [56]. Out of 4000 images with annotated
ground-truth skin regions, 2000 were used to train the global skin color model with the Bayesian clas-
sifier and to determine the DSPF space (there were 1.0 · 108 skin and 4.7 · 108 non-skin pixels in the
training set), and the remaining 2000 images were used as a test set. Here, we used the same setup1 as
for validating our earlier fa-DTCD method [33]. We compare our approach with several state-of-the-art
algorithms listed in Table 1. The experiments were run using a computer equipped with an Intel Core
i7-3740QM 2.7 GHz (16 GB RAM) processor.

1
Available at http://sun.aei.polsl.pl/~mkawulok/icip2013.
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S131

Table 1
Methods investigated during the experimental validation
Abbreviation Reference Description Time (ms)
g-BC [31] Global Bayesian classifier 5.2
g-ESS [14] Global model in the error signal space 4.8
g-RB [15] Global rule-based model in the RGB space 0.8
DTCD [39] Geodesic distance transform in a combined domain 92.9
DSPF [40] Geodesic distance transform using the DSPF space 361.1
CTS [30] Skin detection using color, texture and spatial information 4952.5
fa-ESS [67] Face-based adaptation in the error signal space 24.1
fa-DTCD [33] Geodesic distance transform from face-based adaptive seeds 130.5
sa-DTCD [42] Self-adaptive seeds for the DTCD algorithm 304.6
sa-DSPF [41] Self-adaptive seeds for the DSPF algorithm 548.2
ha-DTCD – Proposed hybrid adaptation for the DTCD algorithm 336.1
ha-DSPF – Proposed hybrid adaptation for the DSPF algorithm 587.4

We evaluate the skin detection accuracy using different measures, including the precision (prec), re-
call (rec), and false positive error rate (δf p ). They depend on the value of the acceptance threshold
used to binarize the SPPMs, and their mutual dependence is rendered in a form of receiver operating
characteristic (ROC) and precision-recall curves. We also quote the area under ROC curve (AUC), the
F-score, and an averaged detection error (δavg = 0.5 · (δf p + (1 − rec))), obtained for the acceptance
threshold set to the values that maximize the F-score and minimize the averaged error. Furthermore, we
verified how much the different methods adapt to the images which do not present the skin at all. To do
so, we removed the skin pixels from the tested images, and we report the false positive rate for these
images (δfNpS ). Naturally, δfNpS = δf p for the pixel-wise schemes.

5.1. Seed extraction

As explained earlier, the seeds should be detected with high precision so as to avoid adapting to the
background, whereas the recall may be low, as it is compensated during the GDT. However, the final
seeds should be located in every ground-truth skin blob to make this compensation possible. So as to
evaluate that, we quote the potential recall (recpot ) – we assume that if a seed pixel is located within
a ground-truth skin region, then the entire region may be potentially correctly detected after spatial
analysis.
In Table 2, we present the skin detection scores obtained at different steps of extracting the seeds.
It may be seen from the table that the scores gradually improve at subsequent steps of the extraction
process for all the investigated approaches. It is worth noting that the potential recall is rather low for
the hybrid adaptation within the initial seeds extracted from facial regions (76.23%), and it is slightly
increased (to 79.76%) after adding the seeds extracted from the entire SPPM. For the self-adaptive
scheme, the potential recall is higher for the initial seeds (94.08%), but the precision is lower than for
the hybrid scheme. It is a noteworthy observation that for the final adapted seeds, the potential recall
in the hybrid-adaptation scheme increases substantially, and after all it exceeds the scores for the self-
adaptive scheme. The precision is increased in both cases, however it is much higher for the hybrid one.
Overall, the F-score is by ca. 2% higher for the hybrid scheme, both for DTCD and DSPF variants.

5.2. Quantitative and qualitative comparisons

The obtained quantitative results of the investigated methods are reported in Table 3. The precision-
recall and ROC curves are presented in Figs 4 and 5, in which our new method is compared with non-
adaptive and adaptive skin detectors, respectively. It may be observed from the plots that ha-DTCD is
S132 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

Table 2
Skin detection scores computed within the seeds
Seed phase ↓ sa-DTCD / sa-DSPF ha-DTCD / ha-DSPF
F-score prec recpot F-score prec recpot
Face seeds – – – 0.8430 94.28% 76.23%
Initial seeds 0.9233 90.64% 94.08% 0.8634 94.11% 79.76%
DTCD Expanded seeds 0.9279 90.10% 95.64% 0.8673 94.80% 79.92%
Adapted seeds 0.9536 91.63% 99.41% 0.9730 95.03% 99.67%
DSPF Expanded seeds 0.9337 91.41% 95.41% 0.8672 94.88% 79.85%
Adapted seeds 0.9539 91.78% 99.29% 0.9738 95.18% 99.68%

Table 3
Skin detection scores obtained using different methods for the ECU data set (bold values indicate the best score)
Acceptance threshold Acceptance threshold Area under
Method set to maximize F-score set to minimize δavg ROC curve
F-score prec rec δfnsp δavg rec δf p δfnsp (AUC)
g-BC 0.7772 73.15% 82.89% 9.13% 12.13% 89.27% 13.52% 13.52% 0.9458
g-ESS 0.7434 68.07% 81.88% 11.79% 14.13% 87.76% 16.03% 16.03% 0.9157
g-RB 0.6896 55.30% 91.61% 23.11% 15.75% 91.61% 23.11% 23.11% 0.8471
CTS 0.7894 76.34% 81.73% 9.01% 12.28% 88.74% 13.31% 13.78% 0.9454
fa-ESS 0.7672 69.67% 85.35% – 13.95% 89.85% 17.74% – 0.9203
DTCD 0.8177 75.79% 88.78% 8.45% 9.87% 92.32% 12.06% 10.32% 0.9338
DSPF 0.8303 78.09% 88.65% 9.06% 7.68% 93.28% 8.64% 12.08% 0.9601
fa-DTCD 0.8661 82.70% 90.92% – 7.17% 94.06% 8.39% – 0.9553
sa-DTCD 0.8543 82.26% 88.85% 10.38% 7.71% 92.18% 7.60% 12.89% 0.9521
sa-DSPF 0.8411 79.10% 89.79% 12.67% 7.22% 94.14% 8.57% 16.57% 0.9609
ha-DTCD 0.8746 85.58% 89.41% 10.96% 6.92% 92.64% 6.47% 12.37% 0.9582
ha-DSPF 0.8542 80.44% 91.05% 13.92% 6.52% 94.61% 7.64% 17.54% 0.9650

(a) (b)

Fig. 4. ROC (a) and precision-recall (b) curves obtained for the ECU data set using non-adaptive methods.

more effective than ha-DSPF, when high precision is required, while for higher recall (i.e., lower value
of the acceptance threshold), the ha-DSPF scheme is more efficient. These two approaches together
dominate the state-of the-art methods for all values of the acceptance threshold, and they are slightly
outperformed by fa-DTCD in terms of the precision, for the recall within the range 93%–95% (see
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S133

(a) (b)

Fig. 5. ROC (a) and precision-recall (b) curves obtained for the ECU data set using adaptive methods.

Fig. 5(b)).
The scores reported in Table 3 clearly indicate that ha-DTCD and ha-DSPF outperform all of the non-
adaptive and adaptive methods in terms of F-score, averaged error δavg and AUC. It can be seen that
AUC and δavg are better using ha-DSPF, while F-score is higher for ha-DTCD. It may also be noticed
that the false positive rate δfNpS for images without skin regions is lower using ha-DTCD than ha-DSPF.
Although for ha-DTCD it is almost twice larger than for the images with skin regions, it is comparable to
that obtained using the global Bayesian classifier (g-BC). Overall, it is clear that the incorrect adaptation
is a potential problem, however we managed to limit its impact using a simple image-level skin detector.
The execution times measured for 512 × 512 images are given in Table 1. The proposed adaptive
methods allow processing ca. 2–3 frames per second (fps), but in case of processing a video stream, the
adaptation may be applied once for a given scene, which should allow processing with similar speeds as
DTCD (10 fps) and DSPF (3 fps).
Several examples of the skin detection outcome are presented in Figs 6 and 7 for non-adaptive and
adaptive methods, respectively. From Fig. 6, it may be seen that ha-DTCD outperforms other methods
in all the cases, apart from the image (IV.), which is better segmented with DSPF. Figure 7 shows that
the hybrid adaptation combines the advantages of face-based adaptation with those offered by the self-
adaptive schemes. For example, the face-based methods fail for the images (II.) and (XIII.) because of
non-skin pixels presence within the facial region, while the hybrid scheme effectively excludes those
pixels from adaptation. The hybrid scheme produces better results also for the images (VII.) and (XIV.),
because it extracts the initial seeds not only from the facial region. The self-adaptive schemes behave
similarly to the hybrid adaptation in the aforementioned cases, but they deliver poor results for the images
(IX.), (XI.) and (XII.), where the hybrid adaptation produces good segmentation outcome. Overall, the
proposed hybrid-adaptive schemes ha-DTCD and ha-DSPF are more stable for the presented set of
images, which has been confirmed quantitatively for the entire test set.

6. Conclusions and future work

In this paper, we introduced a new method for detecting skin seeds for the geodesic distance transform
that combines the benefits offered by our earlier face-based adaptation with those of our recent self-
adaptive algorithm. We have reported the results obtained in our extensive experimental study, which
S134 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

Fig. 6. Examples of skin detection outcome obtained using state-of-the-art non-adaptive methods (red: false positives, blue:
false negatives).
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S135

Fig. 7. Examples of skin detection outcome obtained using adaptive methods (red: false positives, blue: false negatives).
S136 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

clearly indicate that the proposed approach outperforms a number of state-of-the-art skin detectors, and
it delivers better results than our earlier adaptive methods.
The proposed algorithm offers substantial improvement of skin detection accuracy, but still in some
cases the results are below the satisfactory level. This may be addressed by developing higher-level im-
age understanding algorithms to reject the areas which do not present human body, and it is certainly
worth future investigation. Furthermore, the adaptation strategy may involve exploiting information con-
cerned with the phenotype (e.g., ethnicity) – a collection of skin color models could be trained offline,
and each of the models would be adapted to a different phenotypical class, analogously to [49], which
was discussed earlier in Section 3. Here, the phenotype (and corresponding skin color model) would
be selected automatically (e.g., based on face geometry extracted from a grayscale image) or provided
a priori. The selected pre-learned skin color model could be treated as a global model for the method
reported here, thereby combining two different approaches towards skin color model adaptation.
Our ongoing work involves analysis of the developed skin detection methods in terms of their appli-
cability to gesture recognition [25,51]. In particular, we aim at determining, which parts of the operating
curves are most relevant to this application domain, so as to focus our future efforts on that area. This
would also contribute to improving image understanding, as well as the image-level skin detection. The
latter may help reduce the false positive rate for the images that do not present the skin at all.

Acknowledgements

This work was supported by the Polish National Science Center (NCN) under the Grant: DEC-2012/
07/B/ST6/01227.
This work was performed using the infrastructure supported by POIG.02.03.01-24-099/13 grant:
“GeCONiI – Upper Silesian Center for Computational Science and Engineering”.
The authors are grateful to Czestochowa University of Technology for granting access to Intel CPU
and Xeon Phi platforms providing by the MICLAB project No. POIG.02.03.00.24-093/13.
Jolanta Kawulok was supported by the European Union from the European Social Fund (grant agree-
ment number: UDA-POKL.04.01.01-00-106/09).

References

[1] Q. Abbas, M.E. Celebi, I. Fondón García and M. Rashid, Lesion border detection in dermoscopy images using dynamic
programming, Skin Research and Technology 17(1) (2011), 91–100.
[2] A.A. Abin, M. Fotouhi and S. Kasaei, A new dynamic cellular learning automata-based skin detector, Multimedia Sys-
tems 15(5) (2009), 309–323.
[3] S. Alpert, M. Galun, R. Basri and A. Brandt, Image segmentation by probabilistic bottom-up aggregation and cue inte-
gration, in: Proc IEEE Conference on Computer Vision and Pattern Recognition (CVPR ’07) (2007), 1–8.
[4] A.A. Argyros and M.I. Lourakis, Real-time tracking of multiple skin-colored objects with a possibly moving camera,
in: Computer Vision-ECCV 2004, ser. LNCS, T. Pajdla and J. Matas, eds, Springer Berlin Heidelberg 3023 (2004),
pp. 368–379.
[5] H. Baltzakis, M. Pateraki and P. Trahanias, Visual tracking of hands, faces and facial features of multiple persons,
Machine Vision and Applications 23 (2012), 1141–1157.
[6] M. Bello, A combined Markov random field and wave-packet transform-based approach for image segmentation, IEEE
Trans on Image Processing 3(6) (1994), 834–846.
[7] T. Bernas, R. Starosolski, J.P. Robinson and B. Rajwa, Application of detector precision characteristics and histogram
packing for compression of biological fluorescence micrographs, Computer Methods and Programs in Biomedicine
108(2) (2012), 511–523.
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S137

[8] K.K. Bhoyar and O.G. Kakde, Skin color detection model using neural networks and its performance evaluation, Journal
of Computer Science 6(9) (2010), 963–968.
[9] S. Bilal, R. Akmeliawati, M.J.E. Salami and A.A. Shafie, Dynamic approach for real-time skin detection, Journal of
Real-Time Image Processing (2012).
[10] B. Bourdin, Image segmentation with a finite element method, ESAIM: Modélisation Mathématique et Analyse
Numérique 33(2) (1999), 229–244.
[11] T.S. Caetano, S.D. Olabarriaga and D.A.C. Barone, Do mixture models in chromaticity space improve skin detection?
Pattern Recognition 36 (2003), 3019–3021.
[12] H. Castillejos, V. Ponomaryov, L. Nino-de Rivera and V. Golikov, Wavelet transform fuzzy algorithms for dermoscopic
image segmentation, Computational and Mathematical Methods in Medicine 2012 (2012), article ID 578721.
[13] M.E. Celebi, H. Iyatomi, G. Schaefer and W.V. Stoecker, Lesion border detection in dermoscopy images, Computerized
Medical Imaging and Graphics 33(2) (2009), 148–153.
[14] A. Cheddad, J. Condell, K. Curran and P. Mc Kevitt, A skin tone detection algorithm for an adaptive approach to
steganography, Signal Processing 89(12) (2009), 2465–2478.
[15] Y.-H. Chen, K.-T. Hu and S.-J. Ruan, Statistical skin color detection method without color transformation for real-time
surveillance systems, Engineering Applications of Artificial Intelligence 25(7) (2012), 1331–1337.
[16] K. Chenaoua and A. Bouridane, Skin detection using a Markov random field and a new color space, in: Proc IEEE Int
Conf on Image Processing (ICIP 2006) (2006), 2673–2676.
[17] B. Choi, B. Chung and J. Ryou, Adult image detection using Bayesian decision rule weighted by SVM probability, in:
Proc Int Conf on Computer Sciences and Convergence Information Technology (ICCIT ’09) (2009), 659–662.
[18] A. Conci, E. Nunes, J.J. Pantrigo and Á. Sánchez, Comparing color and texture-based algorithms for human skin detec-
tion, in: Proc Int Conf on Enterprise Information Systems (ICEIS 2008) (2008), 166–173.
[19] F. Dadgostar and A. Sarrafzadeh, An adaptive real-time skin detector based on hue thresholding: A comparison on two
motion tracking methods, Pattern Recognition Letters 27(12) (2006), 1342–1352.
[20] J.R. del Solar and R. Verschae, Skin detection using neighborhood information, in: Proc IEEE Int Conf on Automatic
Face and Gesture Recognition (2004), 463–468.
[21] D.A. Forsyth and M.M. Fleck, Automatic detection of human nudes, Int J of Computer Vision 32 (1999), 63–77.
[22] M. Fotouhi, M. Rohban and S. Kasaei, Skin detection using contourlet-based texture analysis, in: Proc Int Conf on
Digital Telecommunications (ICDT ’09) (2009), 59–64.
[23] J. Fritsch, S. Lang, M. Kleinehagenbrock, G. Fink and G. Sagerer, Improving adaptive skin color segmentation by
incorporating results from face detection, in: Proc IEEE Int Workshop on Robot and Human Interactive Communication
(2002), 337–343.
[24] H. Greenspan, J. Goldberger and I. Eshet, Mixture model for face-color modeling and segmentation, Pattern Recognition
Letters 22 (2001), 1525–1536.
[25] T. Grzejszczak, M. Kawulok and A. Galuszka, Hand landmarks detection and localization in color images, Multimedia
Tools and Applications, in print, (2015), 1–25, http://dxdoi.org/10.1007/s11042-015-2934-5.
[26] J. Han, G. Awad, A. Sutherland and H. Wu, Automatic skin segmentation for gesture recognition combining region and
support vector machine active learning, in: Proc IEEE Int Conf on Automatic Face and Gesture Recognition, Washington,
DC, USA: IEEE Computer Society, (2006), 237–242.
[27] L. Ikonen and P. Toivanen, Shortest routes on varying height surfaces using gray-level distance transforms, Image and
Vision Computing 23(2) (2005), 133–141.
[28] B.N. Jagadesh, K. Rao, C. Satyanarayana and G.V.S. RajKumar, Skin colour segmentation using finite bivariate pearso-
nian type-IIb mixture model and k-means, Signal and Image Processing: An Int Journal 3(4) (2012), 37–49.
[29] J. Jaworek-Korjakowska and R. Tadeusiewicz, Assessment of dots and globules in dermoscopic color images as one of
the 7-point check list criteria, in: Proc IEEE Int Conf on Image Processing (ICIP 2013) (2013), 1456–1460.
[30] Z. Jiang, M. Yao and W. Jiang, Skin detection using color, texture and space information, in: Proc Int Conf on Fuzzy
Systems and Knowledge Discovery (FSKD 2007) 3 (2007), 366–370.
[31] M. Jones and J. Rehg, Statistical color models with application to skin detection, Int J of Computer Vision 46 (2002),
81–96.
[32] P. Kakumanu, S. Makrogiannis and N.G. Bourbakis, A survey of skin-color modeling and detection methods, Pattern
Recognition 40(3) (2007), 1106–1122.
[33] M. Kawulok, J. Kawulok, J. Nalepa and M. Papiez, Skin detection using spatial analysis with adaptive seed, in: Proc
IEEE Int Conf on Image Processing (ICIP 2013) (2013), 3720–3724.
[34] M. Kawulok and J. Szymanek, Precise multi-level face detector for advanced analysis of facial images, IET Image
Processing 6(2) (2012), 95–103.
[35] M. Kawulok, Dynamic skin detection in color images for sign language recognition, in: Image and Signal Processing,
ser. LNCS, A. Elmoataz, O. Lezoray, F. Nouboud and D. Mammass, eds, Springer Berlin Heidelberg, 5099 (2008),
pp. 112–119.
S138 M. Kawulok et al. / Hybrid adaptation for detecting skin in color images

[36] M. Kawulok, Energy-based blob analysis for improving precision of skin segmentation, Multimedia Tools and Applica-
tions 49(3) (2010), 463–481.
[37] M. Kawulok, Skin detection using color and distance transform, in: Computer Vision and Graphics, ser. LNCS, L. Bolc,
R. Tadeusiewicz, L. Chmielewski and K. Wojciechowski, eds, Springer Berlin Heidelberg 7594 (2012), 449–456.
[38] M. Kawulok, Texture analysis for skin probability maps refinement, in: Pattern Recognition, ser. LNCS, J.A. Carrasco-
Ochoa, J.F. Martinez-Trinidad, J.A. Olvera Lopez and K.L. Boyer, eds, Springer Berlin Heidelberg 7329 (2012), pp. 75–
84.
[39] M. Kawulok, Fast propagation-based skin regions segmentation in color images, in: Proc IEEE Int Conf on Automatic
Face and Gesture Recognition (2013), 1–7.
[40] M. Kawulok, J. Kawulok and J. Nalepa, Spatial-based skin detection using discriminative skin-presence features, Pattern
Recognition Letters 41 (2014), 3–13.
[41] M. Kawulok, J. Kawulok, J. Nalepa and B. Smolka, Self-adaptive algorithm for segmenting skin regions, EURASIP
Journal on Advances in Signal Processing 2014(170) (2014), 1–22, http://aspeurasipjournals.com/content/2014/1/170.
[42] M. Kawulok, J. Kawulok, J. Nalepa and B. Smolka, Self-adaptive skin segmentation in color images, in: CIARP 2014,
ser. LNCS, E. Bayro-Corrochano and E. Hancock, eds, Springer International Publishing Switzerland, 8827 (2014),
96–103.
[43] M. Kawulok and J. Nalepa, Support vector machines training data selection using a genetic algorithm, in: Structural,
Syntactic and Statistical Pattern Recognition, ser. LNCS, G. Gimel’farb, E. Hancock, A. Imiya, A. Kuijper, M. Kudo, S.
Omachi, T. Windeatt and K. Yamada, eds, Springer Berlin Heidelberg 7626 (2012), pp. 557–565.
[44] M. Kawulok, J. Nalepa and J. Kawulok, Skin detection and segmentation in color images, in: Advances in Low-Level
Color Image Processing, ser. Lecture Notes in Computational Vision and Biomechanics, M.E. Celebi and B. Smolka,
eds, Springer Netherlands 11 (2014), pp. 329–366.
[45] R. Khan, A. Hanbury, R. Sablatnig, J. Stoettinger, F. Khan and F. Khan, Systematic skin segmentation: Merging spatial
and non-spatial data, Multimedia Tools and Applications 69(3) (2014), 717–741.
[46] R. Khan, A. Hanbury, J. Stöttinger and A. Bais, Color based skin classification, Pattern Recognition Letters 33(2) (2012),
157–163.
[47] J. Kovac, P. Peer and F. Solina, Human skin color clustering for face detection, in: Proc Computer As a Tool (EUROCON
2003) 2 (2003), 144–148.
[48] H. Kruppa, M.A. Bauer and B. Schiele, Skin patch detection in real-world images, in: Pattern Recognition, ser. LNCS,
L. Van Gool, ed., Springer Berlin Heidelberg 2449 (2002), 109–116.
[49] J.-S. Lee, Y.-M. Kuo, P.-C. Chung and E.-L. Chen, Naked image detection based on adaptive and extensible skin color
model, Pattern Recognition 40 (2007), 2261–2270.
[50] J. Lichtenauer, M.J.T. Reinders and E.A. Hendriks, A self-calibrating chrominance model applied to skin color detection,
in: Proc Int Conf on Computer Vision Theory and Applications 1 (2007), 115–120.
[51] J. Nalepa, T. Grzejszczak and M. Kawulok, Wrist localization in color images for hand gesture recognition, in: Man-
Machine Interactions 3, ser. Advances in Intelligent Systems and Computing, A. Gruca, T. Czachórski and S. Kozielski,
eds, Springer International Publishing 242 (2014), 79–86.
[52] J. Nalepa and M. Kawulok, Adaptive memetic algorithm enhanced with data geometry analysis to select training data
for SVMs, Neurocomputing 185 (2016), 113–132.
[53] J. Nalepa and M. Kawulok, Adaptive memetic algorithm enhanced with data geometry analysis to select training data
for SVMs, Neurocomputing, in press, (2016), http://dx.doi.org/10.1016/j.neucom2015.12.046.
[54] P. Ng and C.-M. Pun, Skin color segmentation by texture feature extraction and k-mean clustering, in: Int Conf on
Computational Intelligence, Communication Systems and Networks (CICSyN 2011) (2011), 213–218.
[55] S.L. Phung, A. Bouzerdoum and D. Chai, A novel skin color model in ycbcr color space and its application to human
face detection, in: Proc IEEE Int Conf on Image Processing (ICIP 2002) 1 (2002), I-289–I-292.
[56] S.L. Phung, D. Chai and A. Bouzerdoum, Adaptive skin segmentation in color images, in: Proc IEEE Int Conf on
Acoustics, Speech, and Signal Processing (ICASSP ’03) (2003), 353–356.
[57] V.I. Ponomaryov, H. Castillejos and R. Peralta-Fabi, Image segmentation in wavelet transform space implemented on
DSP, in: Proc SPIE, Real-Time Image and Video Processing 2012 8437 (2012), 843 703–843 703–12.
[58] M.-J. Seow, D. Valaparla and V. Asari, Neural network based skin color model for face detection, in: Proc Applied
Imagery Pattern Recognition Workshop (2003), 141–145.
[59] M. Silveira, J. Nascimento, J. Marques, A.R.S. Marcal, T. Mendonca, S. Yamauchi, J. Maeda and J. Rozeira, Comparison
of segmentation methods for melanoma diagnosis in dermoscopy images, IEEE Journal of Selected Topics in Signal
M. Kawulok et al. / Hybrid adaptation for detecting skin in color images S139

Processing 3(1) (2009), 35–45.


[60] K. Sobottka and I. Pitas, Face localization and facial feature extraction based on shape and color information, in: Proc
IEEE Int Conf on Image Processing, ICIP 3 (1996), 483–486.
[61] M. Soriano, B. Martinkauppi, S. Huovinen and M. Laaksonen, Skin detection in video under changing illumination
conditions, in: Proc IEEE Int Conf on Pattern Recognition 1 (2000), 839–842.
[62] R. Starosolski, New simple and efficient color space transformations for lossless image compression, Journal of Vision
Communication and Image Representation 25(5) (2014), 1056–1063.
[63] W.R. Tan, C.S. Chan, P. Yogarajah and J. Condell, A fusion approach for efficient human skin detection, IEEE Trans on
Industrial Informatics 8(1) (2012), 138–147.
[64] A. Taqa and H. Jalab, Increasing the reliability of skin detectors, Scientific Research and Essays 5(17) (2010), 2480–
2490.
[65] X. Wang, X. Zhang and J. Yao, Skin color detection under complex background, in: Proc Int Conf on Mechatronic
Science, Electric Engineering and Computer (MEC 2011) (2011), 1985–1988.
[66] G. Yang, H. Li, L. Zhang and Y. Cao, Research on a skin color detection algorithm based on self-adaptive skin color
model, in: Proc Int Conf on Communications and Intelligence Information Security (ICCIIS 2010) (2010), 266–270.
[67] P. Yogarajah, J. Condell, K. Curran, A. Cheddad and P. McKevitt, A dynamic threshold approach for skin segmentation
in color images, in: Proc IEEE Int Conf on Image Processing (ICIP 2010) (2010), 2225–2228.
[68] B. Zafarifar, A. Martiniere and P. de With, Improved skin segmentation for tv image enhancement, using color and
texture features, in: Proc Int Conf on Consumer Electronics (ICCE 2010) (2010), 373–374.
[69] Q. Zhu, K.-T. Cheng, C.-T. Wu and Y.-L. Wu, Adaptive learning of an accurate skin-color model, in: Proc IEEE Int Conf
on Automatic Face and Gesture Recognition (2004), 37–42.
Copyright of Intelligent Data Analysis is the property of IOS Press and its content may not be
copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for
individual use.

You might also like