Document Image Classification Via AdaBoost and ECOC Strategies
Based on SVM Learners
Mehmet Ahat
1,2
, Cagdas Ulas
1
and Onur Agin
1
1
R&D and Special Projects Department, Yapi Kredi Bank, Gebze, Kocaeli, Turkey
2
Faculty of Engineering and Natural Sciences, Sabanci University, Tuzla, Istanbul, Turkey
Keywords:
Document Image Retrieval and Classification, SVM, One-Versus-All, AdaBoost, ECOC, BoVW Model.
Abstract:
In this paper, we describe easily extractable features and an approach for document image retrieval and clas-
sification at spatial level. The approach is based on the content of the image and utilizing visual similar-
ity, it provides high speed classification of noisy text document images without optical character recognition
(OCR). Our method involves a bag-of-visual words (BoVW) model on the designed descriptors and a Random-
Window (RW) technique to capture the structural relationships of the spatial layout. Using the features based
on these information, we analyze different multiclass classification methods as well as ensemble classifiers
method with Support Vector Machine (SVM) as a base learner. The results demonstrate that the proposed
method for obtaining structural relations is competitive for noisy document image categorization.
1 INTRODUCTION
Document Image Retrieval is a crucial research area
dealing with the problem of retrieving structurally
similar document images from a large heterogeneous
collection given a relevant image, which is useful for
document database management, information extrac-
tion and document routing (Hu et al., 1999). In to-
day’s world, large quantities of paper documents are
converted into electronic form and stored as document
images in digital libraries. Storing scanned document
images alone does not suffice, instead, it is only ben-
eficial when the process of retrieving relevant docu-
ments should be done in an efficient manner. How-
ever, the number of relevant documents provided for
retrieval is usually much lower than the number of ir-
relevant documents, resulting in imbalance problem
in the data and thus, the retrieval problem becomes
more challenging (Zheng et al., 2004).
Several methods have been practised for docu-
ment image retrieval and categorization. Most of
these methods are mainly based on layout (struc-
ture) or content of the documents. Content based ap-
proaches are highly dependent on the quality of OCR
and attaining OCR result on the entire document is
excessively expensive in terms of time (Kumar et al.,
2012). As shown in Figure 1, images from different
types of documents often have quite distinct spatial
layout styles. At very low resolutions, these distinc-
tions are also identifiable which allows us to develop
faster algorithms than that of content based tech-
niques for document classification (Hu et al., 1999).
One of the well known methods on layout similarity
is based on the block segmentation where the image is
divided into several structural blocks(Fan et al., 2001)
and these blocks are then compared to their analogous
for the given type of documents. Another popular
method proposes the creation of spatial-pyramid fea-
tures (Lazebnik et al., 2006) by partitioning the image
into smaller grids and computing the density of fea-
tures in each region. Finding efficient ways to capture
the structural relationships at local level is an impor-
tant research problem, and several methods (Yang and
Newsam, 2011; Kumar and Doermann, 2013) have
been proposed before on this issue.
In this work, we propose a method for the clas-
sification of noisy document images. All the docu-
ments that we consider in classification are extracted
from our real-world bank dataset, consisting various
types of forms (see Figure 1) mostly used in loan
applications and scanned by bank branch employee
to be converted to digital format. Our approach in
this study employs different multiclass classification
methods and AdaBoost based ensemble classifiers
with SVM as a base learner to make predictions on
document type by utilizing a set of features that repre-
sent the structural information at spatial layout level.
Our work differs from previous approaches in sev-
250
Ahat M., Ulas . and Agin O..
Document Image Classification Via AdaBoost and ECOC Strategies Based on SVM Learners.
DOI: 10.5220/0005131502500255
In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2014), pages 250-255
ISBN: 978-989-758-054-3
Copyright
c
 2014 SCITEPRESS (Science and Technology Publications, Lda.)

eral ways: (1) We represent each small region of the
image with a very small number of feature descrip-
tors, which results in high speed for feature extraction
procedure (2) Images are represented with less num-
ber of visual codewords compared with other meth-
ods using SIFT features (Smith and Harvey, 2011)
as the key descriptor (3) Without any use of irrele-
vant images in training phase, the proposed method
can achieve high recall rate for correct detection of ir-
relevant images (4) We compare the performance of
different multiclass classification strategies as well as
ensemble classifiers and demonstrate that the results
are especially competitive when SVM based ensem-
ble classifiers are used for this kind of imbalanced
dataset problem.
The remainder of this paper is structured as fol-
lows: In Section 2, the proposed noisy document im-
age classification method is presented. We provide
a description of the experimental setup and demon-
strate the classification results in Section 3 and finally
we give concluding remarks in Section 4.
2 PROPOSED METHOD
The proposed method in this work is composed of
following steps: the description of feature descrip-
tors, the utilization of BoVW model with Random-
Window approach and the analysis of SVM based
multiclass classification strategies and ensemble clas-
sifiers. Following sections will provide the details of
these steps.
2.1 Feature Descriptors
The document images in our business form database
are binary (monochrome) images that the pixel values
can take only one of two values (0,1), suffering from
lack of information as compared to color and gray-
scale images. Hence, specifically designed features
should be found for this type of images.
In this work, we divide each training and test im-
age into small square patches to determine a 60× 40
image layout with 2400 image patches and represent
each patch with 4 feature descriptors based on struc-
tural variations in small local areas.
If P
C
denotes the width (column number) of the
image patch; P
R
denotes the height (row number) of
the patch and w(i, j) represents the pixel value of the
pixel at ith row and jth column, the feature descriptors
are calculated as follows:
1. Column Standard Deviation (σ
c
)
σ
c
=
p
Variance(X
C
) (1)
Figure 1: Sample business forms in the dataset. Each form
is an example from one of the relevant classes.
where X
C
= [X
1
X
2
...X
P
C
] and each X
i
=
P
R
∑
j=1
w( j,i)
2. Row Standard Deviation (σ
r
)
σ
r
=
p
Variance(X
R
) (2)
where X
R
= [X
1
X
2
...X
P
R
] and each X
i
=
P
C
∑
j=1
w(i, j)
3. Patch Mean Value (m
p
)
m
p
=
1
P
R
P
C
P
C
∑
j=1
P
R
∑
i=1
w(i, j)
!
(3)
4. Pixel Transition Intensity (t
s
)
t
s
=
1
P
R
(P
C
− 1)
P
R
∑
j=1
P
C
∑
i=1
trans( j,i)
!
(4)
where trans( j, i) is defined as follow:
trans( j,i) =
(
1 if w( j,i− 1) = 1 & w( j,i) = 0
0 otherwise
2.2 Bag of Visual Words (BoVW) Model
In computer vision, the BoVW model (Csurka et al.,
2004) can be applied to image classification and re-
lated tasks by treating image descriptors as words. A
bag of visual words is a sparse vector of mostly oc-
currence counts or presence of the visual words from
a vocabulary of local image features. A vocabulary
(or codebook) of visual models is obtained by clus-
tering local image descriptors extracted from training
images, which is also described as vector quantiza-
tion of image features into visual words. The vector
quantization process is generally done by a hard or
soft assignment (clustering) and a codebook of visual
words is obtained. Visual words (codewords) are then
defined as the centers of learned clusters.
DocumentImageClassificationViaAdaBoostandECOCStrategiesBasedonSVMLearners
251

In this work, we use k-means clustering (Winn
et al., 2005) to determine the codebook of visual
words. The number of cluster is empirically set as 4
due to the intuition that the structural variations in-
side the local patches of text-document images are
small. After obtaining the visual words, each train
and test images are represented with a sequence of vi-
sual words in a 60 × 40 layout. This layout is given
as the input to “Random Window” generator where a
pre-defined number of windows inside the layout is
selected and each window is represented with a set of
features based on structural relations.
2.3 Random Window (RW) Approach
As it was mentioned in Section 2.2, we represent our
document images by visual words. However, visual
words are not enough to discriminate the structure of
different documents. Usually each type contains par-
ticular image patterns inside sub-images whose coor-
dinates and sizes are unknown. In order to capture
spatial relationships, after converting all document
images into a sequence of visual words in a 60 × 40
layout, we randomly select rectangular windows in-
side the layout and extract layout features by using
the following approach:
Let RW is the set of randomly selected windows’
coordinates represented as
RW
i
=
n
m
′
i
,m
′′
i

,

n
′
i
,n
′′
i
o
where 0 ≤ m
′
i
≤ m
′′
i
≤ 60 and 0 ≤ n
′
i
≤ n
′′
i
≤ 40.
Let V is the available number of visual words
in vocabulary and S is the set of occurrence counts
of the visual words for given RW’s. Hence, S
i
=

s
i
1
,s
i
2
,...,s
i
V

where s
i
j
is the occurrence count of the
jth visual word in RW
i
. For given RW
i
, the feature
vector F
i
is defined as; F
i
=

s
i
1
/η,s
i
2
/η,...,s
i
V
/η

where η is the normalization constant calculated as
η =
q
(s
i
1
)
2
+ (s
i
2
)
2
+ ... + (s
i
V
)
2
for the set of S
i
.
2.4 Support Vector Machine (SVM)
SVM is basically conceived for binary classification.
The idea is to separate two classes by calculating
the maximum margin hyperplane between the train-
ing examples (Vapnik, 1998). The decision function
of SVM for a binary classification problem is
f(x) = hw,φ(x)i + b (5)
where φ(x) is a mapping of sample x from the input
space to a high dimensional feature space. h.,.i de-
notes the dot product in the feature space. The opti-
mal values of w and b can be determined by solving
the following optimization problem:
minimize g(w,ξ) =
1
2
kwk
2
+C
N
∑
i=1
ξ
i
subject to y
i
(hw,φ(x
i
)i + b) ≥ 1− ξ
i
, ξ
i
≥ 0
(6)
where ξ
i
is the ith slack variable and C is the regular-
ization parameter. The minimization problem in (6)
can be written as
minimize W(α) = −
N
∑
i=1
α
i
+
1
2
N
∑
i=1
N
∑
j=1
y
i
y
j
α
i
α
j
k(x
i
,x
j
)
subject to
N
∑
i=1
y
i
α
i
= 0, ∀i : 0 ≤ α
i
≤ C
(7)
where α
i
is a Lagrange multiplier corresponding to
sample x
i
, k(., .) is a kernel function which implicitly
maps the input vectors into a suitable feature space. In
this space, an optimal separating hyperplane is con-
structed by the support vectors.
k(x
i
,x
j
) = hφ(x
i
),φ(x
j
)i (8)
In this work, we use the RBF kernel, k(x
i
,x
j
) =
exp(−kx
i
− x
j
k)
2
/2σ
2
). The performance of RBF-
SVM is mainly affected by the kernel parameters, for
example, σ, and the regularization parameter, C. By
using model selection techniques such as k-fold or
leave-one-out cross-validation (CV), a single best σ
and C can be found (Li et al., 2008).
The following briefly describes several notations
used in this paper:
• T = {(x
1
,y
1
),(x
2
,y
2
),...,(x
N
,y
N
)} : A training
set; where x
i
∈ R
n
; each label, y
i
is an integer
value belongs to Y = {l
1
,l
2
,...,l
N
c
}, where N
c
is
the number of classes. h = {h
1
,h
2
,...,h
n
} : A set
of n binary classifiers.
2.5 Multiclass SVM Classifiers
2.5.1 One-Versus-All (OVA)
The one-versus-all method constructs n binary clas-
sifiers, one for each class. The ith classifier, h
i
, is
trained with the data from class i as positive instances
and all data from other classes as negative instances
to discriminate among the patterns of the class and
the patterns of the remaining (Bagheri et al., 2012).
A new instance is classified as the class whose corre-
sponding classifier output has the largest value (prob-
ability). Hence, the ensemble decision function, h, is
defined as:
y = argmax
i∈{1,2,...,n}
h
i
(x) (9)
NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications
252

2.5.2 Error Correcting Output Codes (ECOC)
The ECOC framework is widely used for multiclass
classification problems. It is based on combining bi-
nary classifiers and designing a codeword for each
class. Since each class is coded by different code-
words, it may exhibit error-correcting capabilities,
which increases accuracy of the multiclass problem
(Dietterich and Bakiri, 1995). Let M is a coding ma-
trix with a dimension of N
c
× L whose elements m
i, j
can be {−1,+1} for dense ECOC models. L is the
length of codewords which is used for class assign-
ments. Each column of M is a map of binary classi-
fier that separates each class from the others. Later on,
the ECOC method was extended and elements can be
{−1, 0, +1} where “0” labeled elements in the coding
matrix is not considered during the training phase of
a particular binary classifier (Allwein et al., 2001).
M
N
c
,L
=




m
1,1
m
1,2
··· m
1,L
m
2,1
m
2,2
··· m
2,L
.
.
.
.
.
.
.
.
.
.
.
.
m
N
c
,1
m
N
c
,2
··· m
N
c
,L




When an instance x is tested, each binary classifier
predicts -1 or 1 for x and these predictions creates a
L long output code vector. x is labeled as the class
whose codewordhas minimum distance. The distance
is usually hamming distance between the output code
vector and class codewords.
2.6 AdaBoost SVM
Although the use of SVM as a component (base)
classifier in AdaBoost may not seem to be in accor-
dance with Boosting principle since SVM itself is
not a weak classifier, the proposed AdaBoostSVM in
(Li et al., 2008) demonstrates that it can show bet-
ter generalization performance than SVM on imbal-
anced classification problems that we consider in this
study. The key idea of AdaBoostSVM is that for a se-
quence of trained SVM component classifiers, start-
ing with large σ values (implies week learning), the
σ values are reduced progressively as the Boosting
iteration proceeds which effectively produces a set
of component classifiers whose model parameters are
adaptively different resulting in better generalization
as compared to using a fixed (optimal) σ value.
The AdaBoostSVM method (Li et al., 2008) which
was proposed for binary classification problem can be
easily modified for OVA multiclass approach. The
pseudo-code of the OVA-AdaBoostSVM is provided
in Algorithm 1.
Algorithm 1: OVA-AdaBoostSVM.
Input: T, Y, number of training samples, N; the initial
σ, σ
ini
; the minimum σ, σ
min
; the step of σ, σ
step
.
for each l
n
∈ Y do
Apply the class binarization for class l
n
on T.
Initialize: the weights of training samples: w
1
i
=
1
N
.
while σ > σ
min
do
(1) Train a RBF-SVM component classifier, h
t
,
on weighted T.
(2) Calculate the training error of h
t
: ε
t
=
∑
N
i=1
w
t
i
, y
i
6= h
t
(x
i
).
(3) If ε
t
> 0.5, decrease σ value by σ
step
and go
to step (1).
(4) Set the weight of h
t
: α
t
=
1
2
ln(
1−ε
t
ε
t
).
(5) Update the weights of training samples:
w
t+1
i
=
w
t
i
exp{−α
t
y
i
h
t
(x
i
)}
C
t
where C
t
is a normal-
ization constant, and
∑
N
i=1
w
t+1
i
= 1.
end while
f
l
n
(x) = sign(
∑
T
t=1
α
t
h
t
(x))
end for
Output: y(x) = l
n
∈ Y, where f
l
n
(x) = 1.
3 EXPERIMENTAL RESULTS
In this work, we consider the classification problem
of document images from 8 types (classes) of busi-
ness forms as shown in Figure 1. We use our own
data set to evaluate the performance of proposed clas-
sification approach. Our training data consists of 50
image samples for each class. The test data consists of
2066 image samples. 1267 of these images are irrele-
vant images whereas 799 of them are relevant images.
Most of the images in the training and test dataset
were contaminated with marginal (clutter) noise and
salt&pepper noise during scanning, transmission or
conversion to digital form.
Three different methods for classification analy-
sis of this multiclass classification problem are com-
pared in this study: OVA-SVM, OVA-AdaBoost-
SVM, Sparse ECOC-SVM. For OVA and ECOC
based SVM modeling, 5-fold CV conducted for both
parameter tuning and generalization capability. For
AdaBoost-SVM, the regularization parameter C is
empirically set as 10 for all experiments. The σ
min
is
computed as the average minimum distance between
any two training samples inside the subset of training
data and the σ
ini
is set as the L
2
norm of the aver-
age of the training samples in the input space. Lastly,
σ
step
is determined to be 2. For ECOC-SVM, we use
Sparse ECOC models with Hamming Decoding. We
initialize the codebooks (coding matrix) with OVA
class binarization and generate randomly rest of the
codebooks each of whose row (codeword) can have
DocumentImageClassificationViaAdaBoostandECOCStrategiesBasedonSVMLearners
253

Table 1: Performance values for each method when window number is 400. (Irrelevant HR = Irrelevant class avg. hit rate
(recall), Relevant HR = Relevant classes avg. hit rate).
Method Precision Accuracy F
1
score Irrelevant HR Relevant HR
OVA 0.9293 0.9433 0.9176 0.9335 0.9414
AdaBoost 0.9351 0.9520 0.9261 0.9962 0.9144
Sparse ECOC 0.9209 0.9347 0.9062 0.9089 0.9176
zero elements with 0.3 probability. 15 × log(N
c
) bi-
nary SVMs are trained. Moreover, a fresh codebook
is generated with the same settings in each run.
Due to the variations in the results, we run each
method several times using the train and test data. The
final performance of each algorithm on the data set is
the average of the results over all runs. The decision
procedure of each method is as follows: If the pre-
dicted label of a test image is “-1” for all binary clas-
sifiers, then the test image class is assigned to be “0”
class which is also called as irrelevant class. If only
one of the OVA classifiers returns the label “+1”, the
class label of the test image is determined to be the
class holding that OVA classifier. Lastly, If more than
one of the OVA classifiers return “+1”, then the class
satisfying the highest probability estimate or smallest
distance (for ECOC-SVM) is chosen as classification
decision.
The performance evaluation of the proposed algo-
rithm on 3 different SVM based approaches depends
on the following important criteria: precision, recall,
accuracy, F
1
score, irrelevant class hit rate (recall) and
relevant classes hit rate. The macro-weighted aver-
age precision, recall, accuracy and F
1
score values
are calculated as in (Sokolova and Lapalme, 2009).
The irrelevant class hit rate is calculated by dividing
the number of irrelevant images which are correctly
classified by the total number of irrelevant image in
test data using the aforementioned decision proce-
dure. The relevant classes hit rate is determined as
in the same way when assuming that there is not any
irrelevant class and as opposed to aforementioned de-
cision procedure, the class decision for relevant im-
ages is made only according to the highest probability
estimate or smallest distance.
The performance values for each method when
randomly generated window number is fixed to 400
are shown in Table 1. The best results are highlighted
with bold fonts. Results in Table 1 demonstrates that
AdaBoost-SVM method outperforms others on our
own data set in terms of all criteria except for rele-
vant classes hit rate. This method achieves a classi-
fication accuracy of 95.2% and 99.6% irrelevant hit
rate, which almost implies perfect detection of irrel-
evant images. Sparse ECOC-SVM based classifica-
tion method exhibits the worst performance among
all 3 classification approaches with F
1
score = 0.906.
One possible reason of this is that the performance of
ECOC method is drastically affected by the associa-
tion between class label and its codeword representa-
tion (Crammer and Singer, 2000). Generating random
code matrix with fixed settings can create sub-optimal
codeword representations that lead to the lowest per-
formance. Figure 2 shows the average recall values
of all runs versus the number of randomly selected
windows considered as in the range of [50,600]. Gen-
erally, the recall value has an increasing trend with
the number of window due to having more informa-
tion on visual words’ statistics. The best value is
achieved by the case when the window number is 400
and AdaBoost performs the best with 92.8 % recall
rate. Next, the McNemar’s statistical test (Li et al.,
2008) is employed to determine the significance of the
results presented in Table 1. Table 2 shows the Mc-
Nemar’s statistical test results for each method pair.
The results are obtained by averaging the test statis-
tics among all runs for each method pair. The test
results illustrate that the performance of AdaBoost
significantly differs from that of other two methods
on this data set ({7.46,21.69} > 3.8414). The under-
lying reason for this is the Boosting mechanism that
forces several SVM component classifiers on imbal-
anced data sets to focus on the misclassified samples
from the minority class, and to prevent them from be-
ing wrongly classified (Li et al., 2008).
Our results in Table 1 indicates that none of the
three methods can achieve the best performance in
terms of Irrelevant HR and Relevant HR at the same
time. A good future direction of this work can be us-
ing an another strong classifier such as Random De-
cision Forest (RDF) (Yao et al., 2011) combined with
SVM for better prediction of both irrelevant and rele-
vant classes.
Table 2: McNemar’s statistics between all method pairs
when window number is fixed to 400.
Method Pairs McNemar’s statistic (χ
2
)
OVA - AdaBoost 7.46
AdaBoost - Sparse ECOC 21.69
OVA - Sparse ECOC 21.02
NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications
254

0 100 200 300 400 500 600
80
85
90
95
100
Number of randomly selected window
Average Recall (%)
OVA−SVM
OVA−AdaBoost−SVM
Sparse−ECOC−SVM
Figure 2: Average Recall values versus the number of ran-
domly selected window for each method.
4 CONCLUSION
In this paper, we have proposed a method for the clas-
sification and retrieval of business form type docu-
ment images. In our method, we incorporate BoVW
model using a set of features based on structural vari-
ations in local image patches and present an approach
to learn the visual words’ histogram at layout level.
Using a real-world bank data, we perform the analy-
sis of different multiclass classification strategies and
ensemble classifiers (Boosting) method with SVM as
a base learner. Although initial results in this study
seem to be promising, we believe that the proposed
document image classification approach should also
be investigated on real benchmark datasets. Further-
more, the effectiveness of the proposed local feature
descriptors in this work should also be compared with
that of the existing descriptors in literature, e.g., SIFT
and SURF. Both of these issues remain as a future
work to validate the robustness of the proposed ap-
proach.
ACKNOWLEDGEMENTS
This work was partially supported by the Scientific
and Technological Research Council of Turkey under
Grant 3120918 and by Yapi Kredi Bank under Grant
62609.
REFERENCES
Allwein, E., Schapire, R., and Singer, Y. (2001). Reducing
multiclass to binary: A unifying approach for margin
classifiers. J. Mach. Learn. Res., 1:113–141.
Bagheri, M., Montezar, G., and Escalera, S. (2012). Error
correcting output codes for multiclass classification:
Application to two image vision problems. In CSI In-
ternatioanal Symposium on Artificial Intelligence and
Signal Processing (AISP), pages 508–513.
Crammer, K. and Singer, Y. (2000). On the learnability and
design of output codes for multiclass problems. In
Proceedings of the Thirteenth Annual Conference on
Computational Learning Theory, pages 35–46.
Csurka, G., Dance, C., Fan, F., Willamowski, F., and Bray,
C. (2004). Visual categorization with bags of key-
points. In Workshop on Statistical Learning in Com-
puter Vision, ECCV, pages 1–22.
Dietterich, T. and Bakiri, G. (1995). Solving multiclass
learning problems via error-correcting output codes.
J. Artif. Int. Res., 2(1):263–286.
Fan, K., Wang, Y., and Chang, M. (2001). Form docu-
ment identification using line structure based features.
In Proc.of the 6th Int. Conf. on Document Anal. &
Recognition, pages 704–708.
Hu, J., Kashi, R., and Wilfong, G. (1999). Document im-
age layout comparision and classification. In Proc. of
the 6th Int. Conf. on Document Anal. & Recognition,
pages 285–288.
Kumar, J. and Doermann, D. (2013). Unsupervised classi-
fication of structurally similar document images. In
Proc. of the 12th Int. Conf. on Document Anal. &
Recognition, pages 1225–1229.
Kumar, J., Ye, P., and Doermann, D. (2012). Learning Doc-
ument Structure for Retrieval and Classification. In In-
ternational Conference on Pattern Recognition (ICPR
2012), pages 1558–1561.
Lazebnik, S., Schmid, C., and Ponse, J. (2006). Beyond
bags of features: Spatial pyramid matching for recog-
nizing natural scene categories. In Computer Vision
and Pattern Recognition, IEEE Conf, pages 2169–
2178.
Li, X., Wang, L., and Sung, E. (2008). Adaboost with svm-
based component classifiers. Eng. Appl. Artif. Intell.,
21(5):785–795.
Smith, D. and Harvey, R. (2011). Document retrieval using
sift image features. 17(1):3–15.
Sokolova, M. and Lapalme, G. (2009). A systematic anal-
ysis of performance measures for classification tasks.
Inf. Process. Manage., 45(4):427–437.
Vapnik, V. (1998). Statictal Learning Theory. John Wiley
and Sons, Inc.,New York.
Winn, J., Criminisi, A., and Minka, T. (2005). Object cat-
egorization by learned universal visual dictionary. In
ICCV, pages 1800–1807.
Yang, Y. and Newsam, S. (2011). Spatial pyramid co-
occurrence for image classification. In Computer
Vision (ICCV),International Conference on, pages
1465–1472.
Yao, B., Khaslo, A., and Fei-Fei, L. (2011). Combining ran-
domization and discrimination for fine-grained image
categorization. In Proc. CVPR.
Zheng, Z., Wu, X., and Srihari, R. (2004). Feature selection
for text categorization on imbalanced data. SIGKDD
Explor. Newsl., 6(1):80–89.
DocumentImageClassificationViaAdaBoostandECOCStrategiesBasedonSVMLearners
255