Active Contour based Automatic Feedback for Optical Character
Recognition
Joanna Isabelle Olszewska
School of Computing and Technology, University of Gloucestershire, The Park, Cheltenham, GL50 2RH, U.K.
Keywords:
Active Contours, Multi-feature Vector Flow, Tracking, Optical Character Recognition, Pattern Recognition,
Unsupervised Segmentation, Object Detection, Team Sport Video Analysis, Automated Scene Understanding.
Abstract:
In this paper, we present a new optical character recognition approach. Our method combines chromaticity-
based character detection with active contour segmentation in order to robustly extract optical characters from
real-world images and videos. The detected character is recognized using template matching. Our developed
approach has shown excellent results when applied to the automatic identification of team players from online
datasets and is more efficient than the state-of-the-art methods.
1 INTRODUCTION
Automatic scene understanding of team sports (Ol-
szewska and McCluskey, 2011) is essential for sport
events’ refereeing and analysis and it involves vision-
based technologies such as object detection (Alqaisi
et al., 2012), object recognition (Olszewska, 2012a),
tracking (Olszewska, 2012b), or spatio-temporal rea-
soning (Olszewska, 2011).
In particular, automatic identification of team
players is of prime importance to support both
sport comments production and media archiving (Al-
suqayhi and Olszewska, 2013).
For that, face recognition techniques such as
(Wood and Olszewska, 2012) have been applied to
process soccer games. However, this biometric ap-
proach is intrinsically not adapted to identify a player
whose back is turned to the camera, in which case his
face is poorly or not visible at all.
As a result, optical character recognition (OCR)
methods have been developed to recognize numbers
on team player’s uniform. Most of them exploit
the temporal redundancy of a character across sev-
eral frames and thus are only limited to video anal-
ysis (Andrade et al., 2003), (Kokaram et al., 2006),
(D’Orazio and Leo, 2010), (Huang et al., 2006), (Ekin
et al., 2003), (Niu et al., 2008) and not suited for tasks
such as still image dataset retrieval. Other works use
both facial and textual cues (Bertini et al., 2006), but
their computational speed is low.
Hence, in this work, we focus on the sport scene
analysis based on the automatic player identification
in images of any type, relying on the detection and
recognition of the player’s jersey number, and there-
fore, on the development of a full, efficient optical
character recognition (OCR) system for this purpose.
OCR major phases are (i) character extraction and
(ii) character recognition. In the first step, the sys-
tem localizes and extracts the character by detecting
its geometrical features like edges or color features,
or both (Lin and Huang, 2007). In the second step,
character recognition is usually performed by match-
ing (Guanglin and Yali, 2010) or by using classifiers,
e.g. AdaBoost (Chen and Yuille, 2004). However,
these existing OCR systems are mainly applied to rec-
ognize license plate numbers or handwritten charac-
ters, whereas player number recognition presents ad-
ditional challenges. Indeed, the foreground, i.e. the
character, could be highly skewed with respect to the
camera, or the background, i.e. the jersey, could be
folded so that part of the number could be hidden.
Moreover, sport images are often blurred, since cam-
eras or players or both are quickly moving.
In this paper, we propose to automatically extract
characters from images based on their local proper-
ties such as their pixel chromaticity and relying on
their global properties processed by the active con-
tours, while we use a digit template to recognize the
extracted characters, leading to an OCR system ro-
bustly dealing with sport applications, while being
computationally effective.
No temporal redundancy assumption is made in
our method, which is thus valid not only for video
frames, but also for still images such as those con-
318
Olszewska J..
Active Contour based Automatic Feedback for Optical Character Recognition.
DOI: 10.5220/0004935603180324
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (MPBS-2014), pages 318-324
ISBN: 978-989-758-011-6
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Figure 1: Our Optical Character Recognition system’s architecture.
tained in sport datasets or on Internet.
In our approach, players could be identify even in
back profile, since our OCR system detects and recog-
nize characters which could be anywhere on the team
player’s clothes.
Hence, the contribution of this paper is:
• the use of active contours as an automatic feed-
back for the chromatic/achromatic segmentation
approach in order to extract characters robustly;
• the development of a new powerful OCR sys-
tem based on the association of this automatic
feedback for character detection with the template
matching-based technique for the fast character
recognition, in context of the automated identifi-
cation of team players in online image and videos.
The paper is structured as follows. In Section
2, we describe our optical character recognition ap-
proach for fast and effective number extraction and
identification. Our method has been successfully ap-
plied to soccer players’ real-world image datasets as
reported and discussed in Section 3. Conclusions are
presented in Section 4.
2 CHARACTER RECOGNITION
AND IDENTIFICATION
SYSTEM
In this section, we present our optical character recog-
nition approach (Fig. 1) for the reliable identification
of soccer player’s numbers present in real-world im-
ages and videos. Firstly, the studied image is seg-
mented by both chromaticity-based approach and ac-
tive contour approach, as explained in Section 2.1. Fi-
nally, the extracted character is recognized by means
of template matching described in Section 2.2.
2.1 Character Extraction
Character extraction consists here in image segmen-
tation and character detection. On one hand, the im-
age is binarized based on chromaticity properties of
the foreground and background pixels as described
in Section 2.1.1. Next, the characters’ inner bound-
ary tracing algorithm is applied in order to extract the
numbers as presented in Section 2.1.2. On the other
hand, active contours are processed and then, they de-
lineate the boundaries of the character under investi-
gation as explained in Section2.1.3. Hence, this later
approach gives a feedback on the first processed ex-
traction, leading to a more robust character detection.
Algorithm 1 : Achromatic-color Number & Achromatic-
color Jersey
if ((N
S
< Y
S
or N
V
< Y
V
) and (J
S
< Y
S
or J
V
< Y
V
)) then
if J
V
> V
thresh
then
for all P do
if P
V
< V
thresh
then
I
B
(P) = 0 set pixel as black
else
I
B
(P) = 1 set pixel as white
end if
end for
else
for all P do
if P
V
< V
thresh
then
I
B
(P) = 1 set pixel as white
else
I
B
(P) = 0 set pixel as black
end if
end for
end if
end if
return I
B
ActiveContourbasedAutomaticFeedbackforOpticalCharacterRecognition
319
Algorithm 2 : Achromatic-color Number & Chromatic-
color Jersey.
if ((N
S
< Y
S
or N
V
< Y
V
) and (J
S
> Y
S
and J
V
> Y
V
))
then
for all P do
if ((P
S
< Y
S
) and (P
V
< Y
V
)) then
I
B
(P) = 0 set pixel as black
else
if (h
di f f
(J
H
, P
H
) < H
thresh
) then
I
B
(P) = 1 set pixel as white
else
I
B
(P) = 0 set pixel as black
end if
end if
end for
end if
return I
B
Algorithm 3 : Chromatic-color Number & Achromatic-
color Jersey.
if ((J
S
< Y
S
or J
V
< Y
V
) and (N
S
> Y
S
and N
V
> Y
V
))
then
if J
V
> V
thresh
then
for all P do
if ((P
S
< Y
S
) and (P
V
< Y
V
)) then
if P
V
< V
thresh
then
I
B
(P) = 0 set pixel as black
else
I
B
(P) = 1 set pixel as white
end if
else
I
B
(P) = 0 set pixel as black
end if
end for
else
for all P do
if (P
S
< Y
S
and P
V
< Y
V
) then
if P
V
> V
thresh
then
I
B
(P) = 0 set pixel as black
else
I
B
(P) = 1 set pixel as white
end if
else
I
B
(P) = 0 set pixel as black
end if
end for
end if
end if
return I
B
2.1.1 Image Segmentation
Let us consider a color image I, where M and N are
its width and height, respectively. The first step to
extract numbers or foregrounds of this still image is
to separate them from their background. In fact, in
Algorithm 4: Chromatic-color Number & Chromatic-color
Jersey.
if ((N
S
> Y
S
and N
V
> Y
V
) and (J
S
> Y
S
and J
V
> Y
V
))
then
for all P do
if (h
di f f
(N
H
, P
H
) < H
thresh
) then
I
B
(P) = 0 set pixel as black
else
I
B
(P) = 1 set pixel as white
end if
end for
end if
return I
B
football, players’ number color is chosen by the foot-
ball league to be in contrast with players’ kit (shirt and
sweater), in order to allow visibility of the number in
diverse conditions. The study of (Saric et al., 2008)
has found that this contrast is the most important in
the hue, saturation and value (HSV) color space when
looking at the saturation of the number pixels and the
jersey pixels. Consequently, the image I could be seg-
mented based on the low and high saturated pixels,
i.e. objects’ achromatic and chromatic colors, respec-
tively, leading to a binary image I
B
. In particular, a
color pixel under investigation P = [P
H
, P
S
, P
V
] is con-
sidered as achromatic if its saturation (P
S
) is below
the saturation threshold (Y
S
) or if its intensity (P
V
) is
below intensity threshold (Y
V
). If the pixel saturation
and intensity are above these thresholds, then it is con-
sidered as chromatic.
The segmentation is initialized by defining the
mean color vector of the jersey J = [J
H
, J
S
, J
V
] and the
mean color vector for the number N = [N
H
, N
S
, N
V
],
based on provided image samples. Next, the image
I is processed depending if the number color is chro-
matic or achromatic and if the jersey color is chro-
matic or achromatic, leading to four cases, i.e. to four
Algorithms 1-4. The segmentation is based on the hue
threshold H
thresh
and the hue difference in the case of
a chromatic-color jersey, whereas the intensity differ-
ence and the intensity threshold V
thresh
are used in the
case of an achromatic-color jersey (Saric et al., 2008).
In the case where the number has an achromatic color
and the jersey color is chromatic (Algorithm 2), the
hue difference h
di f f
is defined as follows:
h
di f f
(J
H
, P
H
) =
∆(J
H
, P
H
) if∆(J
H
, P
H
) < 180
◦
,
360
◦
− ∆(J
H
, P
H
) otherwise,
(1)
with
∆(J
H
, P
H
) =| J
H
− P
H
| . (2)
When both the jersey and the number have chro-
matic colors, the image is segmented as described in
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
320
Algorithm 4, using the hue difference h
di f f
defined as
follows:
h
di f f
(N
H
, P
H
) =
∆(N
H
, P
H
) if∆(N
H
, P
H
) < 180
◦
,
360
◦
− ∆(N
H
, P
H
) otherwise,
(3)
with
∆(N
H
, P
H
) =| N
H
− P
H
| . (4)
2.1.2 Character Detection
In the binarized image I
B
computed by the process ex-
plained in Section 2.1.1, jerseys appear as white ob-
jects, while numbers as black ones. Based on that
fact, tracing internal boundaries of these objects is
an efficient method for number region localization
and extraction. For this purpose, we have adapted
the Boundary Tracing approach (Ren et al., 2002).
Hence, our process presented in Algorithm 5 initiates
by tracing all the boundaries B
i
within the segmented
binary image, and then, in relation to the specific area
aspect ratio F characterizing the number region, the
boundaries are filtered, in order to select only those
containing numbers. Once this process is completed,
the binary image I
B
is cropped and the cropped image
I
C
is transferred to the recognition stage which then
identifies the numbers as detailed in Section 2.2.
This section has presented the single digit case.
The identification of two-digit numbers is as follows.
If two cropped images are of the same size and are
in adjacent bounding rectangles, they are flagged as
forming a two-digit number.
Algorithm 5: Boundary Tracing.
Step 1
Find boundaries B = {B
i
} of all objects
Step 2
for all B
i
do
if B
i
of black object then
if B
i
dimensions = F dimensions then
x
1
= min(B
i
[1])
y
1
= min(B
i
[2])
x
2
= min(B
i
[1])
y
2
= max(B
i
[2])
I
C
= I
B
[x
1
: x
2
][y
1
: y
2
]
else
Ignore boundary
Go the next boundary
end if
end if
end for
return I
C
2.1.3 Active Contours
In this work, multi-feature vector flow active contours
are used to provide another character segmentation in
order to have a feedback on the results computed in
Sections 2.1.1-2.1.2.
Indeed, multi-feature active contour is a paramet-
ric planar deformable curve C
C
C (s) = [C
C
C
x
(s),C
C
C
y
(s)],
with 0 ≤ s ≤ 1, which evolves from an initial posi-
tion to object boundaries with the use of the innova-
tive MFVF field Ξ
Ξ
Ξ(x, y) = [ξ
u
(x, y), ξ
v
(x, y)].
In this framework, the convergence of the curve
is guided by internal and external forces, which are
involved in a gradient descent process. The internal
forces constrain the active contour shape, in the way
to ensure regularity and smoothness of the curvature.
MFVF external force regroups all the selected fea-
tures in one original bidirectional force, enabling the
active contour to reach its final accurate position, even
in complex situations.
Formally, the deformable curve C
C
C (s,t) is modeled
itself by a B-spline paradigm in order to be computa-
tionally efficient, and must satisfy the following dy-
namic equations,
C
C
C
xt
(s,t) = α C
C
C
00
x
(s,t) − β C
C
C
0000
x
(s,t) + ξ
u
(x, y) (5)
C
C
C
yt
(s,t) = α C
C
C
00
y
(s,t) − β C
C
C
0000
y
(s,t) + ξ
v
(x, y), (6)
where C
C
C
00
x
, C
C
C
00
y
, C
C
C
0000
x
, C
C
C
0000
y
, respectively, are the sec-
ond and fourth-order derivatives with respect to the
parameter s of the curve, α is the curvature elasticity
coefficient, and β is the curvature rigidity coefficient.
The active contour, found by solving (5) and (6),
could be, in practice, roughly initialized from a dis-
tance of the target, as the MFVF force offers a large
capture range. This obtained fast multi-feature ac-
tive contour owns high-deformation capabilities that
are well suited for tracking non-rigid objects whose
shapes change markedly. Indeed, tracking with the
multi-feature active contour could be performed by
minimizing the associated energy functional, for each
corresponding feature, in each frame.
Moreover, this computed curve enables precise
foreground segmentation, without any kind of as-
sumption about the object appearance (Olszewska,
2012b).
2.2 Character Recognition
For the recognition of the characters extracted either
with the chromaticity-based technique or active con-
tour method, we have adopted template matching ap-
proach. Indeed, this pattern classification method
is well suited in the identification of small regions
(Brunelli, 2009), which is the case in our application.
ActiveContourbasedAutomaticFeedbackforOpticalCharacterRecognition
321
(a) (b) (c) (d)
(e) (f)
(g)
(h)
Figure 2: Examples of results obtained with our OCR system. First column: input image. Second column: chromaticity-based
segmentation. Third column: active contour-based segmentation. Fourth column: recognized character.
The basis of template matching is that a processed
image is compared to each of the images stored within
a template. In many instances, the extracted number
region has smaller or larger dimensions compared to
the template dimensions, or has not the same orienta-
tion. Thus, the extracted number image has first to be
rotated and rescaled to fit the template orientation and
size, respectively. Then, the correlation coefficient r
between the two compared images is computed as fol-
lows:
r =
∑
m
∑
n
(T
mn
−
¯
T )(S
mn
−
¯
S)
q
[
∑
m
∑
n
(T
mn
−
¯
T )
2
]
∑
m
∑
n
(S
mn
−
¯
S)
2
, (7)
where T
mn
are the values of the pixels of the tem-
plate image with an m ×n size and a mean
¯
T ; S
mn
are
the values of the pixels of the processed image, i.e.
the rescaled cropped binarized image, with a mean
¯
S.
When the structure of the processed image is
greatly similar to the structure of one of the template
images, then the correlation coefficient value is high
and this means the number is identified.
A digit is considered as recognized when at least
one segmentation technique has succeed to extract it
and when the template matching has processed suc-
cessfully and coherently. In case when the two types
of segmentation followed each by template matching
provide different results, the digit is flagged as unrec-
ognized.
To recognize two-digit numbers, single numbers
flagged as constituting a two-digit number in Section
2.1 are recognized individually by matching each of
them with the template. The two-digit number is then
formed based on that information.
We can notice that the use of the template match-
ing technique is well suited for our system of au-
tomatic number recognition of soccer players. On
one hand, template matching is particularly fast when
used in context of our system, because it requires only
the recognition of numerical characters, rather than a
wider range of alphanumerical characters as in other
applications, such as license plate recognition (LPR).
Indeed, our template stores in total only 10 images of
one-digit numbers (0 to 9). Hence, the matching is
performed against a maximum of ten stored images,
in order to recognize the extracted character, which
is computationally very efficient. Moreover, the scale
sensitivity of the template matching technique is used
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
322
Table 1: Average rates of the automatic character extraction and the automatic character recognition obtained for all the
dataset using approaches of 3(Bertini et al., 2006), 2(Saric et al., 2008), 4 (Alsuqayhi and Olszewska, 2013), and our.
Rate 3 2 4 our
character extraction rate 80.0% 83.0% 88.0% 95.0%
character recognition rate 67.5% 52.0% 86.0% 90.0%
in our work as an advantage, since smaller dimensions
of the template dimensions lead to a faster matching.
On the other hand, the recognition rate obtained by
our implementation of this method in our system is
much higher than those presented in the literature as
discussed in Section 3.
3 RESULTS EVALUATION AND
DISCUSSION
To validate our method, we have carried out exper-
iments which consist in automatically recognizing
numbers from the soccer players’ jerseys within a
database containing data images with soccer-related
content, as such illustrated in Fig. 2.
For this purpose, our system has been applied on
a dataset containing 4500 football images whose av-
erage resolution is of 230x330 pixels and which were
captured in outdoor environment. This database owns
challenges of quantity, pose and scale variations of the
players. Moreover, the colors of the teams’ uniforms
have various colors and the fonts on the players’ jer-
seys could vary strongly.
All the experiments have been run on a computer
with an Intel(R) Core(TM)2 Duo 2.53 GHz proces-
sor, 4 Gb RAM, and using our OCR software imple-
mented with MatLab. Our system is able to support
different types of image formats such as jpeg, tiff,
bmp, and png.
In order to assess the performance of our OCR
system, we use the following criteria:
extraction rate =
CL × 100
T T
, (8)
recognition rate =
CR × 100
T T
, (9)
with CL, the number of correctly localized char-
acters, CR, the number of correctly recognized char-
acters, and T T , the total number of tested characters.
Some examples of the results of our OCR sys-
tem are presented in Fig. 2. These samples present
difficult situations such as variability of the jerseys’
colors, i.e. different pixels’ chromaticity properties
of the foregrounds and the backgrounds; numbers’
changing characteristics, i.e. different characters’ ge-
ometrical and spatial properties; scale effects such as
zoom out or close-up.
We can observe that using our approach, charac-
ters are correctly extracted and correctly recognized,
despite their geometrical and chromatical differences.
Hence, our OCR system is robust towards changes in
numbers and colors of the foregrounds and the back-
grounds as well as towards variations of fonts, size,
and orientation of the characters. Moreover, the sys-
tem is robust even in case the chromatic detection pro-
vides a sparse result such as in Fig. 2(f), because of
the effect of the feedback provided by the active con-
tours as displayed in Fig. 2(e).
In Table 1, we have reported the extraction and
recognition rates of our OCR method against the rates
achieved by approaches using chromatic/achromatic
segmentation (C/A Segm.) or template matching (TM)
MSERE + TM (Bertini et al., 2006), C/A Segm. + CL
(Saric et al., 2008), and (C/A Segm. + TM) (Alsuqayhi
and Olszewska, 2013).
We can see in Table 1 that our OCR method rely-
ing on the active contour feedback into the OCR pro-
cess which combines chromatic/achromatic segmen-
tation and matching-based recognition outperforms
the state-of-art approaches for soccer player’s num-
ber identification. In particular, we can notice than
the extraction rate is improved when using the ac-
tive contour as feedback for the chromatic/achromatic
segmentation instead of using C/A segm. alone. Our
OCR method outperforms also other state-of-the-art
techniques such as maximally stable extremal region
extraction. On the other hand, we can observe the
positive effect of our active contour based automatic
feedback approach on the recognition rate compared
to other classification methods.
From Table 1, we can conclude also that the incor-
poration of the active contours increases the robust-
ness of the OCR system. Indeed, it helps in improving
the detection rate, thus the recognition rate is higher
as well.
Moreover for all the dataset, the average compu-
tational speed of our combined OCR method is in the
range of few seconds, and thus, our developed system
could be used in context of online scene analysis.
ActiveContourbasedAutomaticFeedbackforOpticalCharacterRecognition
323
4 CONCLUSIONS
Reliable team player identification in online data such
as images and videos is a challenging topic we have
copped with. For this purpose, we have developed
a new OCR approach relying on both chromaticity-
based segmentation and active contour method which
provides a feedback to the system to reinforce the ro-
bustness of the character extraction. Template match-
ing is used for the character recognition step. Our
OCR system shows greater performance than the ones
found in the literature in both extraction and recogni-
tion of soccer players’ numbers. Moreover, our OCR
approach is well suited for the automatic retrieval and
analysis of online, visual data about team sports.
REFERENCES
Alqaisi, T., Gledhill, D., and Olszewska, J. I. (2012). Em-
bedded double matching of local descriptors for a fast
automatic recognition of real-world objects. In Pro-
ceedings of the IEEE International Conference on Im-
age Processing (ICIP’12), pages 2385–2388.
Alsuqayhi, A. and Olszewska, J. I. (2013). Embedded dou-
ble matching of local descriptors for a fast automatic
recognition of real-world objects. In Proceedings
of the IAPR International Conference on Computer
Analysis of Images and Patterns Workshop (CAIP’13),
pages 139–150.
Andrade, E. L., Khan, E., Woods, J. C., and Ghanbari,
M. (2003). Player classification in interactive sport
scenes using prior information region space analy-
sis and number recognition. In Proceedings of the
IEEE International Conference on Image Processing
(ICIP’03), pages III.129–III.132.
Bertini, M., Bimbo, A. D., and Nunziati, W. (2006). Match-
ing faces with textual cues in soccer videos. In Pro-
ceedings of the IEEE International Conference on
Multimedia and Expo, pages 537–540.
Brunelli, R. (2009). Template Matching Techniques in Com-
puter Vision: Theory and Practice. John Wiley and
Sons.
Chen, X. and Yuille, A. L. (2004). Detecting and read-
ing text in natural scenes. In Proceedings of the
IEEE Computer Society Conference on Computer Vi-
sion and Pattern Recognition, pages II.366–II.373.
D’Orazio, T. and Leo, M. (2010). A review of vision-based
systems for soccer video analysis. Pattern Recogni-
tion, 43(8):2911–2926.
Ekin, A., Tekalp, M., and Mehrotra, R. (2003). Auto-
matic soccer video analysis and summarization. IEEE
Transactions on Image Processing, 12(7):796–806.
Guanglin, H. and Yali, G. (2010). A simple and fast method
of recognizing license plate number. In Proceedings
of the IEEE International Forum on Information Tech-
nology and Applications, pages II.23–II.26.
Huang, C.-L., Shih, H.-C., and Chao, C.-Y. (2006).
Semantic analysis of soccer video using dynamic
Bayesian network. IEEE Transactions on Multimedia,
8(4):749–760.
Kokaram, A., Rea, N., Dahyot, R., Tekalp, A. M.,
Bouthemy, P., Gros, P., and Sezan, I. (2006). Brows-
ing sports video: Trends in sports-related indexing and
retrieval work. IEEE Signal Processing Magazine,
23(2):47–58.
Lin, C.-C. and Huang, W.-H. (2007). Locating license plate
based on edge features of intensity and saturation.
In Proceedings of the IEEE International Conference
on Innovative Computing, Information and Control,
pages 227–230.
Niu, Z., Gao, X., Tao, D., and Li, X. (2008). Semantic
video shot segmentation based on color ratio feature
and SVM. In Proceedings of the IEEE International
Conference on Cyberworlds, pages 157–162.
Olszewska, J. I. (2011). Spatio-temporal visual ontology.
In Proceedings of the EPSRC Workshop on Vision and
Language.
Olszewska, J. I. (2012a). A new approach for automatic ob-
ject labeling. In Proceedings of the EPSRC Workshop
on Vision and Language.
Olszewska, J. I. (2012b). Multi-target parametric active
contours to support ontological domain representa-
tion. In Proceedings of the RFIA Conference, pages
779–784.
Olszewska, J. I. and McCluskey, T. L. (2011). Ontology-
coupled active contours for dynamic video scene un-
derstanding. In Proceedings of the IEEE International
Conference on Intelligent Engineering Systems, pages
369–374.
Ren, M., Yang, J., and Sun, H. (2002). Tracing boundary
contours in a binary image. Image and Vision Com-
puting, 20(2):125–131.
Saric, M., Dujmic, H., Papic, V., and Rozic, N. (2008).
Player number localization and recognition in soccer
video using HSV color space and internal contours. In
Proceedings of the World Academy of Science, Engi-
neering and Technology, pages 531–535.
Wood, R. and Olszewska, J. I. (2012). Lighting-variable
AdaBoost based-on system for robust face detection.
In Proceedings of the International Conference on
Bio-Inspired Systems and Signal Processing, pages
494–497.
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
324
