Expert Systems With Applications 42 (2015) 8989–9000

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Improving the descriptors extracted from the co-occurrence matrix

using preprocessing approaches

Loris Nanni a,∗, Sheryl Brahnam b, Stefano Ghidoni a, Emanuele Menegatti a

a Department of Information Engineering, University of Padua, Via Gradenigo, 6, 35131 Padova, Italyb Department of Computer Information Systems, Missouri State University, 901 S. National, Springfield, MO 65804, USA

a r t i c l e i n f o

Keywords:

Co-occurrence matrix

Texture descriptors

Preprocessing

Local approaches

Multi-scale

a b s t r a c t

In this paper, we investigate the effects that different preprocessing techniques have on the performance of

features extracted from Haralick’s co-occurrence matrix, one of the best known methods for analyzing image

texture. In addition, we compare and combine different strategies for extracting descriptors from the co-

occurrence matrix. We propose an ensemble of different preprocessing methods, where, for each descriptor,

a given Support Vector Machine (SVM) classifier is trained. The set of classifiers is then combined by weighted

sum rule. The best result is obtained by combining the extracted descriptors using the following preprocess-

ing methods: wavelet decomposition, local phase quantization, orientation, and the Weber law descriptor.

Texture descriptors are extracted from the entire co-occurrence matrix, as well as from sub-windows, and

evaluated at multiple scales. We validate our approach on eleven image datasets representing different im-

age classification problems using the Wilcoxon signed rank test. Results show that our approach improves the

performance of standard methods. All source code for the approaches tested in this paper will be available

at: https://www.dei.unipd.it/node/2357

© 2015 Elsevier Ltd. All rights reserved.

1

m

d

s

d

p

n

p

S

s

i

t

p

u

S

m

(

R

O

a

R

E

(

P

e

C

C

(

a

P

s

I

a

i

h

0

. Introduction

Medical imaging is experiencing rapid technological develop-

ents and, as a result, is producing unprecedented amounts of

ata that are now being warehoused in specialized medical re-

earch databases around the world1. It is generally agreed that these

atabases contain information that, if appropriately harnessed, could

otentially revolutionize scientific knowledge in medicine. What is

eeded to make this dream a reality is the development of more so-

histicated and robust machine vision algorithms.

Texture analysis is essential for many image classification tasks.

ome of the best-performing methods rely on features that encap-

ulate specific information about the elementary characteristics

nherent in a given input image. A feature extractor algorithm ob-

ains a vector of numbers, called a “descriptor”, from an image that

rovides a specific description of that image. Features that are widely

sed in the computer vision and robotics communities include the

cale-Invariant Feature Transform (SIFT) (Lowe, 2004), Speeded Up

∗ Corresponding author. Tel.: +39 3493511673.

E-mail addresses: nanni@dei.unipd.it, loris.nanni@unibo.it (L. Nanni), sbrahnam@

issouristate.edu (S. Brahnam), ghidoni@dei.unipd.it (S. Ghidoni), emg@dei.unipd.it

E. Menegatti).1 For instance, Biograph, mGen, PathogenPortal, and ConsensusPathDB

L

n

F

h

b

p

ttp://dx.doi.org/10.1016/j.eswa.2015.07.055

957-4174/© 2015 Elsevier Ltd. All rights reserved.

obust Feature (SURF) (Bay, Tuytelaars, & Gool, 2006), Histogram of

riented Gradients (HOG) (Dalal & Triggs, 2005), Gradient Location

nd Orientation Histogram (GLOH) (Mikolajczyk & Schmid, 2005),

egion Covariance Matrix (RCM) (Tuzel, Porikli, & Meer, 2006),

dgelet (Wu & Nevatia, 2005), Gray Level Co-occurrence Matrix

GLCM) (Haralick, Shanmugam, & Dinstein, 1973), Local Binary

atterns (LBP) (Ojala, Pietikäinen, & Harwood, 1996), non-binary

ncodings (Paci, Nanni, Lathi, Aalto-Setälä, & Hyttinen, 2013), Color

orreloGram (CCG) (Huang, Kumar, Mitra, Zhu, & Zabih, 1997), Color

oherence Vectors (CCV) (Pass, Zabih, & Miller, 1996), color indexing

Swain & Ballard, 1991), steerable filters (Freeman & Adelson, 1991),

nd Gabor filters (Jain & Farrokhnia, 1991).

An especially powerful feature extractor is the Local Binary

attern (LBP) operator (Ojala, Pietikainen, & Maeenpaa, 2002), which

tands out because of its simplicity, effectiveness, and robustness.

t has proven to be a powerful discriminator in a number of im-

ge classification problems, e.g., in discriminating endoscopical

mages of pediatric celiac diseases (Vécsei, Amann, Hegenbart,

iedlgruber, & Uhl, 2011) and in distinguishing real masses from

ormal parenchyma in mammographic images (Oliver, Lladó,

reixenet, & Martí, 2007). LBP is also useful in data mining, where it

as been combined with other descriptors, for instance, to classify

rain magnetic resonance data (Unay & Ekin, 2008) and different cell

henotypes (Nanni & Lumini, 2008).

8990 L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000

p

O

u

l

c

e

s

t

w

p

f

a

i

c

d

e

d

e

m

a

a

a

e

c

w

b

o

I

c

e

o

w

e

t

t

a

e

w

t

a

2

w

H

d

This paper is a continuation of our research group’s previous

explorations (Nanni, Brahnam, Ghidoni, & Menegatti, 2013d; Nanni,

Brahnam, Ghidoni, Menegatti, & Barrier, 2013b; Nanni, Ghidoni,

& Menegatti, 2013c; Nanni, Paci, Brahnam, Ghidoni, & Menegatti,

2013a) for improving one of the first methods for analyzing texture,

the GLCM, which was first proposed by Haralick, (1979) for analyzing

satellite images. The GLCM is one of the most studied and most

extensively used texture analysis techniques: its application ranges

from visual inspection for industrial production (Zhao et al., 2014) to

analysis of microscopy imagery (Pantic et al., 2014). An alternative

version of the co-occurrence matrix capable of dealing with color

is proposed in Subrahmanyam, Wu, Maheshwari, and Balasubra-

manian (2013); the formulation they offer takes into account the

inter-correlation between the different color channels of a RGB

(Red-Green-Blue) color image.

The GLCM is usually analyzed by means of a set of descriptors,

commonly referred to as Haralick descriptors, that are evaluated by

means of a process that provides a single value for each single de-

scriptor starting from the whole matrix (as it happens when mean

and variance are evaluated from a probability distribution). Within

the last few years, the GLCM has become the focus of several re-

search groups whose aim is to discover novel methods for extract-

ing yet more information from the GLCM. Some important work in

this area includes that of Gelzinis, Verikas, & Bacauskiene (2007),

where different values of the distance parameter influencing the

GLCM are considered; (Walker, Jackway, & Longstaff, 2003), where

features are extracted from areas presenting high discrimination us-

ing the weighted summation of GLCM elements; (Chen, Chengdong,

Chen, & Tan, 2009), where descriptors are obtained from the gray

level-gradient co-occurrence matrix by calculating the gradient value

of each pixel in the neighborhood of interest points; and (Mitrea,

Mitrea, Nedevschi, Badea, & Lupsor, 2012), where the GLCM is com-

bined with the edge orientation co-occurrence matrix of superior or-

der (see, Akono, Tankam, & Tonye, 2005) by taking into considera-

tion the gray levels of the image pixels and some local features such

as edges. In addition, multi-scale analysis has been performed using

the GLCM, where descriptors are extracted from each rescaling opera-

tion. For instance, in Rakwatin, Longepe, Isoguchi, Shimada, and Uryu

(2010) the image is rescaled to different sizes, and in Hu, (2009) and

Pacifici and Chini (2008) multiple scales are investigated by chang-

ing the window size used to extract GLCM descriptors. In some very

recent work, LBP has been combined with the GLCM. In Sastry, Ku-

mari, Rao, Mallika, and Lakshminarayana (2012), for instance, the

GLCM is constructed after applying LBP to the image, and features

are extracted from second-order statistical parameters from the re-

sulting matrix. In Sun, Wang, Chen, She, and Kong (2012), GLCM is

constructed from an LBP image produced after the application of a

Gaussian filtering pyramid preprocessing step.

A major difficulty encountered when analyzing texture is the

strong dependency some texture analysis methods have on image

resolution and scale. Several researchers have recently reported the

effectiveness of combining more than one multi-scale approach

along with LBP descriptors to overcome this problem (see, for in-

stance, Ren, Wang, & Zhao, 2012). In Nanni et al. (2013c) and Nanni

et al. (2013d), our group applied this idea to the co-occurrence

matrix, demonstrating that results can be enhanced when different

descriptors extracted from the co-occurrence matrix are combined

with a multi-scale approach.

In addition to combine the GLCM with texture descriptors and the

multi-scale approach, our main intention in this paper is to com-

pare and to combine (for the first time) the application of differ-

ent preprocessing techniques to each image before calculating the

co-occurrence matrix. An extensive study comparing and combining

preprocessing methods with the GLCM is important in order to clar-

ify which combinations of methods offer the best performance gain

using a system based on the co-occurrence matrix. The different pre-

rocessing methods explored in this paper are detailed in Section 3.

ur best result is obtained by combining extracted GLCM descriptors

sing the following preprocessing methods: wavelet decomposition,

ocal phase quantization, orientation, and the Weber law descriptor.

As in our previous work (Nanni et al., 2013b) (which did not in-

lude preprocessing methods), we utilize the following methods for

xtracting features:

• The standard approach proposed by Haralick (1979);• A set of 3D features that consider the co-occurrence matrix as a

3D shape (Nanni et al., 2013c);• Gray-level run-length features (Tang, 1998), that consider groups

of consecutive pixels with the same gray level, and measure their

lengths. In this paper, the gray-level run-length features (GL) are

evaluated on the GLCM, and it is important to distinguish GL from

the GLCM. As noted in Section 2.1, the GLCMs are calculated at

different values of d (the distance between the two pixels that are

compared) and θ (the direction they are aligned). As described in

more detail in Section 2.4, the GL approach has its own orientation

(labelled θGL). For each GLCM, the GL descriptor is extracted for all

values of θGL.

We also examine for the first time a novel approach for repre-

enting the co-occurrence matrix using Spatial Diversity (SD) fea-

ures (Geraldo Junior, da Rocha, Gattass, Silva, & de Paiva, 2013). SD

orks poorly when applied to the whole co-occurrence matrix, but

erforms quite well when applied to a set of sub-windows extracted

rom the co-occurrence matrix, i.e., the GLCM considered only on

subset of its 256 × 256 domain. This is an important point, as

t demonstrates that multiple descriptors, if appropriately applied,

an be useful in describing the co-occurrence matrix: this, in turn,

emonstrates our idea that the co-occurrence matrix can be consid-

red as a texture image. Hence, it is possible to apply standard texture

escriptors (e.g., SD, GL, …) to the GLCM to extract information.

In this study we also improve the performance of the feature

xtraction algorithms by extracting features from the co-occurrence

atrix applied to the original image as well as to the image after

gaussian filtering at multiple scales. The co-occurrence matrix is

lso divided into different sub-windows, from which feature sets

re also extracted. Separate SVM classifiers are trained, one for

ach sub-window. The overall classification outcome is obtained by

ombining the separate SVMs decisions using either a sum rule or a

eighted sum rule.

In Nanni et al. (2013b) only the standard approach proposed

y Haralick (1979) and a set of 3D features considering the co-

ccurrence matrix as a 3D shape are used for describing the GLCM.

n this work we test four different approaches for describing the

o-occurrence matrix, adding as well the innovation of combining

ach approach with a set of preprocessing methods. Moreover, in

ur previous work, our approach was validated only on five datasets,

hereas in this study we employ a set of eleven datasets.

The remainder of this paper is organized as follows. The feature

xtractors used in our experiments are described in Section 2, and

he preprocessing methods (and resulting processed images) are de-

ailed in Section 3. The eleven datasets used to evaluate our approach

nd our rationale for using them are presented in Section 4, and the

xperimental results are reported and discussed in Section 5. Finally,

e conclude in Section 6 with some remarks on the limitations of

his study; we also highlight the importance of studying the GLCM

nd offer some suggestions for future investigations.

. Feature extractors

As mentioned in the introduction, in Nanni et al. (2013c, 2013d),

e showed how it is possible to enhance the performance of

aralick’s features by applying the multi-scale approach. We also

emonstrated the effectiveness of combining features extracted from

L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000 8991

d

m

G

t

p

i

u

fi

a

d

w

m

w

o

h

2

m

M

e

t

i

F

t

p

l

r

p

(

a

p

e

t

t

m

s

a

2

f

f

1

1

1

1

m

d

e

Fig. 1. Illustration of the co-occurrence matrix as a 3D function. A set of planes is in-

tersected with the GLCM to obtain a set of level curves that are analyzed to measure

the shape of the GLCM.

2

c

i

m

z

t

t

d

s

t

o

o

a

s

o

g

e

e

L

h

2 Here we report only the best configuration to reduce the size of the tables report-

ing experimental results.3 The weight of the SVMs trained using features extracted from the original image is

4; the other SVMs have a weight of 1.

ifferent sub-windows of the co-occurrence matrix. In brief, the

ulti-scale approach generates images by applying a 2D symmetric

aussian lowpass filter of kernel size k (we use k = 3 and k = 5 in

his work) with standard deviation 1.

Described in this section are each of the feature extraction ap-

roaches used in the experiments reported in Section 5. In all exper-

ments, SVM (Vapnik & Chervonenkis, 1964; Vapnik & Lerner, 1963),

sing both linear and radial basis function kernels, is the base classi-

er. The best kernel and the best set of parameters for each dataset

re chosen using a 5-fold cross-validation approach on the training

ata. The parameters of the fusions are chosen based on criteria that

ill be detailed below and that were empirically discovered to maxi-

ize the performance obtained by the descriptors extracted from the

hole image and from the sub-windows. It should be noted that in

ur experiments the same weights are always used without any ad-

oc dataset tuning.

.1. The gray level dependency matrix (GLCM)

The Gray Level Co-occurrence Matrix (GLCM) (Haralick, 1979) is a

atrix that is obtained as a histogram on a 2D domain of dimension

×M, where M is the number of gray levels in the image (in our case

qual to 256). The GLCM counts the number of pixel transitions be-

ween two pixel values. In other words, the bin of the histogram is

ncremented whose indices are equal to the values of the two pixels.

or example, a pixel transition between two adjacent pixels from 20

o 30 will increment the GLCM matrix at indices (20, 30). Since the

ixels of the input image take values in the range (0, 255), the gray

evel dependency matrix has size 256×256; each bin at indices (also

eferred to as coordinates) (i, j) represents the number of pixel cou-

les showing the transition from gray level i to gray level j.

The way pixel couples are created depends on two parameters: d

the distance between the two pixels) and θ (the direction they are

ligned). A value of d = 1 and θ = 0, for instance, would result in

ixel couples that are adjacent to each other on the same row. In our

xperiments, we consider four directions: the horizontal (H), the ver-

ical (V), the diagonal top left-bottom right, or right-down (RD), and

he top right-bottom left, or left-down (LD).

In this study the GLCM is processed both by considering the whole

atrix, as well as by restricting to some subsets of its domain (called

ub-window): this means that only a subset of the pixel transitions

re considered at a time.

.2. The standard Haralick statistics (HAR)

Haralick (1979) was the first to consider the idea of using a set of

eatures to extract information from the co-occurrence matrix. The

ollowing Haralick features are considered in our tests:

1. Energy

2. Correlation

3. Inertia

4. Entropy

5. Inverse difference moment

6. Sum average

7. Sum variance

8. Sum entropy

9. Difference average

0. Difference variance

1. Difference entropy

2. Information measure of correlation 1

3. Information measure of correlation 2

In HAR a set of 13 features is calculated from each co-occurrence

atrix evaluated at θ = {0°, 45°, 90°, 135°} and d = {1, 3}. A global

escriptor is obtained by concatenating the descriptors obtained for

ach distance and orientation value.

We also evaluated the performance of the following HAR variants:

• HR, in which features are extracted from the entire image only;• HRsub, in which a feature set is extracted from the entire co-

occurrence matrix plus from four sub-windows defined2 using

the coordinates (0, 0) to (127, 127); (128, 128) to (255, 255),

(0, 128) to (255, 128), and (0, 128) to (128, 255); these are the

coordinates that provided the best experimental results in Nanni

et al. (2013d). Each set of extracted features is then used to train

a separate SVM, with the results combined by weighted sum

rule. The weights are 1 for each SVM running on the feature

evaluated on a sub-window, and 4 for the SVM analyzing the

data extracted from the whole image: this choice reflects the fact

that each sub-window has an area that is ¼ the area of the entire

co-occurrence matrix;• HRsca, in which features are extracted from the original image

and two filtered images. In other words, the feature sets are ex-

tracted from the entire co-occurrence matrix and from the four

sub-windows defined in HRsub as well as from the original image

and from the two filtered images. In this way 15 sets are obtained,

and the 15 SVMs are combined by weighted sum rule3.

.3. Shape

This approach explores the shape of the co-occurrence matrix by

onsidering it as a 3D function. SHAPE has been described in detail

n Ghidoni, Cielniak, and Menegatti (2012); Nanni et al. (2013d). The

ain idea behind SHAPE is to intersect the GLCM with a set of hori-

ontal planes defined by some given height values (see Fig. 1) to ob-

ain a set of level curves (which are curves that describe the shape of

he 3D GLCM at given height values). These level curves are used to

erive a set of features.

Each level curve can be composed of more than one blob. The

hape shown in Fig. 1 is the ideal configuration, but spikes can occur

owards the edge of the histogram. To filter these spikes out so that

nly the main component that is representative of the overall texture

f the image is considered, the main blob (the one with the largest

rea) is selected for the feature extraction process. To ease the analy-

is, it is approximated by fitting it to an ellipse (with only a small loss

f information).

Starting at the base of the co-occurrence matrix with height 1 and

oing up by a distance of 2, level curves are progressively consid-

red until height 19. The maximum height value of 19 is based on

xperimental observation of the effects of the different level curves.

evel curves that have the most information lie at a relatively low

eight because the lower regions of the co-occurrence matrix are

8992 L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000

w

e

w

r

o

m

S

G

w

t

t

t

i

w

very stable, unlike the upper regions, which are less stable due to im-

age noise. To understand the impact of noise in the upper regions,

consider input images of similar texture content, say two consecutive

frames of a video in which no perceivable changes occur; the level

curves evaluated on the two images, or video frames, would show

more significant difference when in fact there is very little difference

in the two images because of the noise in the higher level curves.

This is an unwanted influence of the noise affecting the image pixel.

The exact size chosen for the height values also depends on the im-

age size. The values mentioned above were chosen for images of size

640×480, but they are suitable also for smaller and larger images, as

shown in the experimental section. Finally, it should be mentioned

that the co-occurrence matrix is not normalized because normaliza-

tion to the highest bin would introduce instabilities. Certainly, it is

the case that other types of normalization could be performed with

respect to the total volume of the co-occurrence matrix; however, this

volume is actually the total number of pixel couples in the image and

depends on the image dimensions, which is constant for a given video

flow. Normalizing using a constant would just scale the values, mak-

ing the normalization irrelevant.

A set of features extracted from the ellipses obtained as described

above is evaluated at each level. The features describing all levels are

then analyzed together, and a set of nine features are extracted that

describe the evolution of the level curves (see Ghidoni et al., 2012,

and Nanni et al., 2013d, for details). These features provide a charac-

terization of the input image that can be directly used as the input to

a classifier.

The features in SHAPE, as defined in this paper, are evaluated on

the entire co-occurrence matrix as well as on 12 sub-windows (#1-

#12) of the GLCM, which are defined by the following index ranges:

#1: (0, 0) to (127, 127), #2: (128, 128) to (255, 255), #3: (0, 0) to (191,

191), #4: (64, 64) to (255, 255), #5: (0, 0) to (95, 95), #6: (31, 31)

to (95, 95), #7: (63, 63) to (127, 127), #8: (95, 95) to (159, 159), #9:

(127, 127) to (191, 191), #10: (159, 159) to (223, 223), #11: (191, 191)

to (255, 255), and #12: (63, 63) to (191, 191). In SHAPE we extract

more sub-windows than the four previously used because, differently

from that case, the performance of SHAPE improves when smaller

sub-windows are used, as described in (Nanni et al., 2013c). The di-

vision of sub-windows follows the same process carried out in Nanni

et al. (2013c), which proved to be the most effective solution when

dealing with image classification based on Haralick’s features.

For each of these 13 windows (counting the whole GLCM and the

12 sub-windows), a different feature vector is extracted and used as

the input for a separate SVM. Results of the 13 SVMs are combined by

weighted sum rule. A weight of 1 is given to the first five descriptors

and a weight of 0.5 is assigned to the remainder.

Each set of 13 descriptors is derived from the co-occurrence ma-

trices evaluated at θ = {0°, 45°, 90°, 135°} and d = {1, 3} (in the case

where multiple values are used for the distance). The feature vec-

tor is the concatenation of the features extracted for each value of

the distance. In the experimental section, SH denotes the case where

features are extracted only from the entire co-occurrence matrix.

SHsub denotes the method that extracts features from all 13 windows.

SHsca3 is the approach that combines SHsub with the multi-scale

approach.

2.4. Gray-level run-length features (GL)

GL (Tang, 1998) derives features from a run-length matrix that is

calculated using the characteristics of gray level runs (a set of con-

secutive pixels with the same value). The run length is determined

by the number of pixels in the set. Accordingly, the run-length matrix

P contains in each location P(i, j) the number of runs of length j of a

given gray level i. The run length can be evaluated following four dif-

ferent orientations (horizontal, vertical and diagonal), described by

means of the angle θ = {0◦, 45◦, 90◦, 135◦}.

GL

In this work, the run-length matrix is used in a slightly different

ay: instead of measuring the pixel gray levels of a given image, it is

mployed to measure the height of the bins of the GLCM: in other

ords, the GLCM is considered by the run-length evaluation algo-

ithm like an image of size 256×256 whose “pixels” (namely the bins

f the histogram) can have a value above 256.

The run-length matrix is a tool capable of evaluating the unifor-

ity of the images, that is, the areas showing the same gray level.

uch tool is employed in this study to measure this aspect of the

LCM histograms: the run-level matrix is a complementary aspect

ith respect to the SHAPE features described in Section 2.3: while

he former is focused on the uniform regions, the latter is focused on

he variations of the GLCM peaks.

It is possible to evaluate several features from the run-length ma-

rix; in our experiments, we consider the following ones, described

n (Tang, 1998):

• Short Run Emphasis (SRE), evaluated as:

SRE = 1

nr

M∑i=1

N∑j=1

P(i, j)

j2(1)

• where nr is the total number of runs, M the number of gray levels,

and N the maximum run length;• Long Run Emphasis (LRE):

LRE = 1

nr

M∑i=1

N∑j=1

P(i, j) · j2 (2)

• Gray Level Nonuniformity (GLN):

GLN = 1

nr

M∑i=1

(N∑

j=1

P(i, j)

)2

(3)

• Run Length Nonniformity (RLN):

RLN = 1

nr

N∑j=1

(M∑

i=1

P(i, j)

)2

(4)

• Run Percentage (RP):

RP = nr

np(5)

here np is the total number of pixels in the image;

• Low Grey-Level Run Emphasis (LGRE):

LGRE = 1

nr

M∑i=1

N∑j=1

P(i, j)

i2(6)

• High Grey-Level Run Emphasis (HGRE):

HGRE = 1

nr

M∑i=1

N∑j=1

P(i, j) · i2 (7)

• Short Run Low Grey Level Emphasis (SRLGE):

SRLGE = 1

nr

M∑i=1

N∑j=1

P(i, j)

i2 · j2(8)

• Short Run High Grey Level Emphasis (SRHGE):

SRHGE = 1

nr

M∑i=1

N∑j=1

P(i, j) · i2

j2(9)

• Long Run Low Grey Level Emphasis (LRLGE):

LRLGE = 1

nr

M∑i=1

N∑j=1

P(i, j) · j2

i2(10)

L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000 8993

A

p

p

G

a

d

o

a

e

t

G

o

u

e

i

i

p

G

2

o

a

h

c

b

g

e

t

e

t

s

i

i

w

d

t

p

i

H

T

w

M

T

w

H

T

a

T

T

D

T

q

B

A

f

l

H

i

H

T

i

B

F

w

f

E

W

3

i

i

c

a

(

m

m

v

t

p

e

(

r

p

4 In order to better visualize the processed images (PI), the PI are resized to the size

of the original image.

• Long Run High Grey Level Emphasis (LRHGE):

LRHGE = 1

nr

M∑i=1

N∑j=1

P(i, j) · i2 · j2 (11)

ll the above features are evaluated for all the four orientations of θGL

reviously described.

In this study, the run-length matrix is not used to analyze the in-

ut image, as it is often the case, instead it is used to analyze the

LCM (which is a matrix, and can therefore be analyzed using im-

ge processing techniques). In particular, we exploit the descriptors

efined in Eqs. (1)–(11) from a run-length matrix that is evaluated

n the GLCM. As noted in Section 2.1, several GLCMs are calculated

t different values of d = {1, 3} (the distance between the two pix-

ls) and θ = {0°, 45°, 90°, 135°} (the direction they are aligned). To

ake into account all the combinations of such parameters, a single

L descriptor concatenates all the described features (1)–(11). A set

f GLCMs is then evaluated using the four values of θ and the two val-

es of d, listed in Section 2.2, for a total of eight GLCMs. For each of the

ight GLCMs a given GL feature set is extracted. The GL approach has

n turn its own set of orientations: θGL = {0°, 45°, 90°, 135°}, which

ncludes all the values originally considered in Tang (1998). In the ex-

erimental section, we also explore the performance of the following

L variants:

• GR, as in HR but using the GL features;• GRsub, as in HRsub but using the GL features;• GRsca, as in HRsca but using the GL features.

.5. Spatial diversity (SD)

The concept of diversity comes from the field of ecology and is one

f its central themes. Diversity represents a measure of how species

re present in the environment and, more specifically, in different

abitats and communities. The analysis of diversity considers con-

epts like population, diffusion, and diversification and is performed

y means of indices that analyze the presence of a species within a

iven environment.

To apply diversity analysis to pattern recognition, it is first nec-

ssary to express the data to be analyzed (in this case, images) in

erms of population, species, and related concepts. In Geraldo Junior

t al. (2013), the population P is defined as the set of all Np pixels in

he image, while the gray level values constitute the different species

(i); the total number of species is S, which can be smaller than 256

f some gray level values are not present in the image. The number of

ndividuals of a species s(i) is the number n(i) of pixels in the image

ith a given gray level. Finally, the relative frequency for a species is

efined as: p(i) = n(i)/Np. Given this model it is possible to exploit

he indices commonly used to measure population diversity and ap-

ly them to pattern recognition.

The richness of a population is defined as the Shannon–Wiener

ndex, which is defined as:

= −S−1∑i=0

p(i) ln p(i) (12)

his value increases with higher heterogeneity levels.

Another measure of diversity is provided by the McIntosh index,

hich is defined as:

c = Np −√∑S−1

i=0 n(i)2

Np −√

Np

(13)

he Brillouin index, yet another measure of diversity, is best suited

hen the randomness of the population is not guaranteed:

b = 1

Np

(ln (Np!) −

S−1∑i=0

ln (n(i)!)

)(14)

he total diversity index measures species variation, and is defined

s:

d =S−1∑i=0

1

n(i)(p(i)(1 − p(i))) (15)

he Simpson index is a second order statistics, and is defined as:

s =∑S−1

i=0 n(i)(n(i) − 1)

Np(Np − 1)(16)

he Berger–Parker index is the relative frequency of the most fre-

uent species:

p = maxi

p(i). (17)

dditional indices can be defined based on those above. The J index,

or example, compares the value of the H index for the current popu-

ation to the same index provided by a maximally diverse population:′: J = H/H′, with H′ = ln S. The same concept applied to the Simpson

ndex leads to the Ed index: d = Ds/Ds′ .

The Hill index is defined as:

ill =1

Ds−1

eH − 1(18)

he Buzas–Gibson index measures the uniformity of the Shannon

ndex:

g = eH

S(19)

inally, the last index considered in this work is the Camargo index,

hich compares the relative frequencies of the different species as

ollows:

= 1 −(

S−1∑i=0

S−1∑j=i+1

p(i) − p( j)

S

)(20)

e also explore the performance of the following SD variants:

• SD, as in HR but using the SD features;• SDsub, as in HRsub but using the SD features;• SDsca, as in HRsca but using the SD features.

. Preprocessing methods

The primary aim of this work is to improve performance us-

ng different methods for preprocessing the images before extract-

ng features from the co-occurrence matrix. Before calculating the

o-occurrence matrix, however, we normalize the preprocessed im-

ge I to [0, 255], such that each element becomes I’ = [(I – min) ./

max –min)].×255, where max and min represent the maximum and

inimum values found in the images used for training, respectively;

oreover, .× and ./ represent the per-element multiplication and di-

ision, respectively.

In Fig. 2,4 we provide some example images derived by applying

he preprocessing techniques mentioned above. After the set of pre-

rocessed images is available, a set of descriptors is extracted from

ach one: this set is then fed into an SVM. The outputs of the SVMs

one for each preprocessing technique) are finally combined by sum

ule. Below we briefly describe all the preprocessing methods ex-

lored in this work.

8994 L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000

Fig. 2. Examples of preprocessing of a given input image (acronyms are explained be-

low). For WAV we report the approximation coefficients matrix.

,

3

p

l

p

L

w

t

e

t

n

n

a

m

t

b

s

T

c

t

b

T

L

I

a

a

e

3

e

e

c

t

d

a

∇

T

(

c

T

3.1. Wavelet (WAV)

A common operation used in many vision problems is the Wavelet

transform. WAV (Mallat, 1989) performs a single-level 2-D wavelet

decomposition, which requires a 2-D scaling function, ζ (x, y), and

three 2-D wavelets, ψ i(x, y), where the index i = {H, V, D} is a super-

script representing intensity variations along columns, or horizon-

tal edges (ψH), along rows, or vertical edges (ψV), and along diago-

nals (ψD). Given the separable scaling function and the three wavelet

functions, we first define the scaled and translated basis functions as

follows:

ζ j,s1,s2(x, y) = 2 j/2ζ(2 jx − s1, 2 jy − s2

),

ψ ij,s1,s2(x, y) = 2 j/2ψ i

(2 jx − s1, 2 jy − s2

), i = {H,V, D} (21)

The discrete wavelet transform of function f(x, y) of size S1 × S2 is

then:

Wζ ( j0, s1, s2) = 1√S1×S2

S1−1∑x=0

S2−1∑y=0

f (x, y)ζ j0,s1,s2(x, y),

W iψ( j, s1, s2) = 1√

S1×S2

S1−1∑x=0

S2−1∑y=0

f (x, y)ψ ij,s1,s2(x, y) i = {H,V, D}

(22)

where j0 is an arbitrary starting scale and the Wζ (j0, s1, s2) coeffi-

cients define the approximation of f(x, y) at scale j0. The W iψ( j, s1, s2)

coefficients add horizontal, vertical, and diagonal details for scales j

≥ j0.

In our experiments, we use the Daubechies wavelet with four van-

ishing moments, single scale decomposition (i.e. j = 1 and j0 = 1). The

four preprocessing images are the approximation coefficient matrix

and the three detail coefficient matrices (horizontal, vertical, and di-

agonal). The preprocessed images are thus Wζ and W iψ

, i = {H,V, D}.

Notice that the processed images have a size that differs from that of

the original image. Let us define sx × sy the size of the original image

and lf the length of the wavelet filter (here lf = 7) then the size of the

processed image is floor((sx + lf − 1)/2) × floor((sx + lf − 1)/2).

.2. Local binary patterns (LBP)

The canonical LBP operator (Ojala et al., 2002) is computed at each

ixel location of an image by considering the values of a small circu-

ar neighborhood (with radius R pixels) around the value of a central

ixel Ic:

BP(N, R) =N−1∑n=0

s(In − Ic)2n (23)

here N is the number of pixels in the circular neighborhood, R is

he radius, In are the pixels in the current neighborhood (the varying

lement in the sum) and s(x) = 1 if x ≥ 0, otherwise s(x) = 0.

LBP has proven to be highly discriminative and invariant to mono-

onic gray level changes. Unfortunately, it is not very robust to the

oise present in images when the gray-level changes result from

oise that is not monotonic (Tan & Triggs, 2007). Therefore, we use

well-known variant based on ternary coding (Paci et al., 2013) that

akes it more robust to noise. The ternary coding is achieved by in-

roducing the threshold τ (in this work τ = 3) so that the s(x) function

ecomes:

(x) =

⎧⎨⎩

1, x ≥ τ

0, |x| ≤ τ

−1, x ≤ −τ

(24)

he ternary code obtained by the s(x) function is split into two binary

odes by considering its positive and negative components according

o the following binary function bv(s(x)):

v(s(x)) ={

1 s(x) = v0 otherwise

v ∈ {−1, 1}. (25)

hus, the LTPv operator for v∈{−1, 1} is:

TPv(N, R) =N−1∑n=0

bv(s(In − Ic))2n. (26)

n this work we use (R = 1; N = 8) and (R = 2, N = 16). Four codes

re produced for each pixel (the two sets of parameters and, for both,

positive and a negative LBP code), resulting in four images (one for

ach code). The size of the processed image is (sx - R) × (sy – R).

.3. Gradients (GRA)

Gradient is a common image operation typically used to detect

dges. A gradient magnitude image is defined as an image in which

ach pixel measures the change in intensity in a given direction of a

orresponding point in the original image. GRA calculates the magni-

ude of the gradients of each pixel in the x (horizontal) and y (vertical)

irections based on its neighbors.

For a function f(x, y) the gradient of f at coordinates (x, y) is defined

s the 2D column vector:

F =[

Gx

Gy

]=

⎡⎢⎣

∂ f

∂x

∂ f

∂y

⎤⎥⎦ (27)

he processed image used for extracting the co-occurrence matrix at

x, y) is given by the magnitude of the gradient vector at the same

oordinates and is defined as:

∇ f = mag(∇F) =[Gx

2 + Gy2]1/2 =

[(∂ f

∂x

)2

+(

∂ f

∂y

)2]1/2

.

(28)

he size of the processed image is the same of the original image.

L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000 8995

3

o

u

e

θ

w

t

3

u

o

i

e

2

p

o

t

(

l

o

i

n

d

v

w

t

o

e

G

T

d

ξ

W

t

t

3

o

i

a

a

s

H

f

v

i

t

p

∠

E

t

m

a

S

n

i

F

w

2

s

u

l

H

I

W

T

m

t

H

b

G

g

c

t

t

q

w

c

c

b

T

p

u

fi

c

3

u

n

t

t

w

G

m

.4. Orientation (OR)

In OR the image gradient is soft quantized (Vu, 2013) using three

rientations, which result in three processed images. This method is

sed to reduce the noise and other forms of degradation in an image.

Given the gray level of a pixel p(i) at (x, y) in the gradient image

valuated for the orientation of p(i):

(x, y) = atan2(Gy, Gx), (29)

here θ(x, y) = θ(x, y) + π in the case where θ (x, y) ≥ 0.

OR is the magnitude of p(i) distributed over n discretized orienta-

ions as detailed in Vu, (2013).

The size of the processed image is the same of the original image.

.5. Weber law descriptor (WLD)

WLD is based on Weber’s law which states that a change of stim-

lus that is just noticeable for human beings is a constant ratio of the

riginal stimulus; if the change is less than this constant ratio then

t is considered background noise. WLD has two components: differ-

ntial excitation and orientation (Chen, Shan, He, Zhao, & Pietikäinen,

010). For a given pixel, the differential excitation component is com-

uted based on the ratio between (1) the relative intensity differences

f a current pixel compared to its neighbors and (2) the intensity of

he current pixel. The neighborhood is defined as a square of size LNE

we use LNE = 3). From the differential excitation component, both the

ocal salient patterns in the input image and the gradient orientation

f the current pixel is computed. We use the differential excitation

mage for extracting the co-occurrence matrix, and adopt the same

otation proposed in (Chen et al., 2010).

To compute a differential excitation ξ (Ic) of a current pixel Ic, the

ifferences between its neighbors must be calculated:

00s =

L2NE−1∑i=0

(Ii − Ic), (30)

here Ii(i = 0, 1, . . . L2NE − 1) defines the ith neighbors of Ic and L2

NE is

he number of neighbors. The ratio of the differences to the intensity

f the current point of both the original image v01s and the differential

xcitation image v00s is calculated by:

ratio(Ic) = v00s

v01s

. (31)

he arctangent function is then employed on Gratio(∗) to calculate the

ifferential excitation ξ (Ic):

(Ic) = arctan

[v00

s

v01s

]= arctan

[L2

NE−1∑i=0

(Ii − Ic

Ic

)]. (32)

hen ξ (Ic) is positive, it simulates surroundings that are lighter than

he current pixel; when ξ (Ic) is negative, it simulates surroundings

hat are darker than the current pixel.

The size of the processed image is (sx − 1) × (sy − 1).

.6. Local phase quantization (LPQ)

LPQ (Ojansivu & Heikkila, 2008) is based on the blur invariance

f the Fourier Transform Phase. The spatially invariant blurring of an

mage f(x) results in the blurred image, g(x), which can be expressed

s g(x) = f(x)∗h(x), where x = [x,y]T is the spatial coordinate vector

nd h(x) is the point spread function. What this means in the Fourier

pace can be expressed as G(u) = F(u) · H(u), where G(u), F(u), and

(u) are the Discrete Fourier Transforms (DFT) of the blurred g(x),

(x) and h(x), respectively, and u = [u, v]T is the frequency coordinate

ector. The magnitude and phase aspects can be separated resulting

n |G(u)| = |F(u)| · |H(u)| and �G(u) = �F(u) + �H(u). The Fourier

ransform is always real-valued if h(x) is centrally symmetric, and its

hase is a two-valued function given by

H(u) ={

0, H(u) ≥ 0π, H(u) < 0

(33)

q. (32) means that �G(u) = �F(u) for all H(u) ≥ 0; in other words,

he phase of the observed image �G(u) is invariant to centrally sym-

etric blur at the frequencies where H(u) ≥ 0. LPQ is blur invari-

nt because it uses the local phase information obtained from the

hort Time Fourier Transform (STFT) computed over a rectangular

eighborhood Nx of size LNE×LNE at each pixel position x of the

mage f(x):

(u, x) =∑y∈Nx

f (x − y)e− j2πuT y = wuT f x (34)

here y is a spatial coordinate vector; wu is the basis vector of the

-D DFT at frequency u, and fx is a vector containing all the image

amples from Nx.

LPQ considers four frequency vectors: u1 = [a,0]T, u2 = [0, a]T,

3 = [a, a]T, and u4 = [a, a]T, where a is small enough to reside be-

ow the first zero crossing of H(u) that satisfies �G(u) = �F(u) for all

(u) ≥ 0.

If we let Fcx = [F(u1,x), F(u2,x), F(u3,x), F(u4,x)] and Fx = [Re{Fc

x},

m{Fcx}]T, then the corresponding transform matrix is:

= [Re{wu1, wu2, wu3, wu4,}, Im{wu1, wu2, wu3, wu4, }]T

hus, Fx = W fx.

The coefficients need to be decorrelated before quantization to

aximally preserve the information; after decorrelation the samples

o be quantized are statistically independent. For details, see Rahtu,

eikkilä, Ojansivu, and Ahonen (2012). Assuming a Gaussian distri-

ution, a whitening transform can achieve independence, such that

x = VTFx, where V is an orthonormal matrix derived from the sin-

ular value decomposition of the covariance matrix of the transform

oefficient vector Fx. After computing Gx for all the image positions,

he resulting vectors are quantized using the following scalar quan-

izer:

j ={

1, Gj ≥ 00, Gj ≤ 0

, (36)

here Gj represent the jth component of Gx. These quantized coeffi-

ients are represented as integers between 0 and 255 using the binary

oding

=8∑

j=1

qj2j−1 (37)

he image that contains the integers obtained with Eq. (36) is the

reprocessed image. In this work, two sizes for the local window are

sed (LNE = 3 and LNE = 5), both with Gaussian derivative quadrature

lter pairs for local frequency estimation. In this way, two prepro-

essed images are obtained.

The size of the processed image is (sx + 1 − LNE) × (sy +1 − LNE).

.7. Minimum paths by Dijkstra’s algorithm (DJ)

DJ (De Mesquita Sá Junior, Backes, & Cortez, 2013) models images

sing a graph where each pixel corresponds to a node that is con-

ected to its 4-neighbors. The cost, we(x, y), of a transition between

wo connected nodes at locations (x, y) and (x′, y′) is determined by

heir gray levels I(x, y) and I(x′, y′), such that:

e(x, y) = |I(x, y) − I(x′, y′)| + I(x, y) + I(x′, y′)2

(38)

iven this model, the image texture is measured by evaluating the

inimum cost for connecting four point pairs: (i) the central points

8996 L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000

Table 1

Descriptive summary of the datasets.

Name Abreviation #Classes #Samples Sample size Download

Histopatology HI 4 2828 Various http://www.informed.unal.edu.co/histologyDS

Pap smear PAP 2 917 Various http://labs.fme.aegean.gr/decision/downloads

Virus types classification VIR 15 1500 41 × 41 http://www.cb.uu.se/∼gustaf/virustexture

Breast cancer BC 2 584 various upon request to Geraldo Braz Junior [ge.braz@gmail.com]

Protein classification PR 2 349 various upon request to Loris Nanni [nanni@dei.unipd.it]

Chinese Hamster Ovary CHO 5 327 512 × 382 http://ome.grc.nia.nih.gov/iicbu2008/hela/index.html#cho

2D HeLa HE 10 862 512 × 382 http://ome.grc.nia.nih.gov/iicbu2008/hela/index.html

Locate Endegenous LO 10 502 768 × 512 http://locate.imb.uq.edu.au/downloads.shtml

Brodatz BR 13 2080 205 × 205 http://dismac.dii.unipg.it/rico/data.html

Pictures PI 13 903 Various http://imagelab.ing.unimo.it/files/bible_dataset.zip

Painting PA 13 2338 Various http://www.cat.uab.cat/∼joost/painting91.html

p

d

d

e

o

f

e

m

d

t

b

m

e

c

d

f

t

L

2

c

5

d

of the first and last column (p0); (ii) the lower left and upper right

corner (p45); (iii) the central points of the upper and lower row (p90);

and (iv) the lower right and upper left corner (p135). To find the min-

imum paths, Dijkstra’s algorithm (Dijkstra, 1959) is exploited since

link costs cannot be negative.

The cost w depends on two elements: the difference in the gray

levels of the two pixels and their average value. The cost of the four

paths (i–iv described above) will then be higher if a large number of

strong transitions is present in the paths; this links the path costs to

image texture.

In our experimental section, we divide each image into patches of

4 × 4 pixels. For each patch the four minimum cost paths are calcu-

lated. For each of the four paths (i–iv), a different image is built where

the cost path of patch (i,j) of the original image becomes the (i,j) ele-

ment of the new processed image. Since we have four different cost

paths, we process the image in four different ways. Notice that, since

we use patches of 4 × 4 pixels, the size of the processed image is 25%

of the size of the original image (one pixel for each patch).

4. Datasets

To assess our system, we validate it across eleven image datasets,

representing several different computer vision problems. Table 1 pro-

vides a short descriptive summary of each of the following datasets:

• PAP: the pap smear dataset in Jantzen, Norup, Dounias, and

Bjerregaard (2005), containing images representing cells used in

cervical cancer diagnosis.• VIR: the virus dataset in Kylberg, Uppström, and Sintorn (2011),

containing images of viruses extracted by negative stain transmis-

sion electron microscopy. We use the same 10-fold validation di-

vision of images, which we obtained from the authors of Kylberg

et al. (2011). However, rather than use the mask provided in VIR

for background subtraction, we use the entire image for extracting

features since we found this method performs better.• HI: the histopatology dataset in Cruz-Roa, Caicedo, and González

(2011), containing images taken from different organs that repre-

sent the four fundamental tissues.• BC: the breast cancer dataset in Junior, Cardoso, de Paiva, Silva,

and Muniz de Oliveira (2009), containing 273 malignant and 311

benign breast cancer images.• PR: the protein dataset in Nanni, Shi, Brahnam, and Lumini (2010),

containing 118 DNA-binding Proteins and 231 Non-DNA-binding

proteins, with texture descriptors extracted from the 2D distance

matrix that represents each protein. The 2D matrix is obtained

from the 3D tertiary structure of a given protein by considering

only atoms that belong to the protein backbone (see Nanni et al.,

2010, for details).• CHO: the Chinese hamster ovary cell dataset in Boland, Markey,

and Murphy (1998), containing 327 fluorescent microscopy im-

ages of ovary cells belonging to five different classes. Images are

16 bit grayscale of size 512 × 382 pixels.

• HE: the 2D HeLa dataset (Boland & Murphy, 2001), containing 862

single cell images divided into 10 staining classes taken from flu-

orescence microscope acquisitions on HeLa cells.• LO: the LOCATE ENDOGENOUS mouse sub-cellular organelles

dataset (Fink et al., 2006), containing 502 images unevenly dis-

tributed among 10 classes of endogenous proteins or features of

specific organelles.• BR: the exact same set of textures used in Gonzalez, Fernandez,

and Bianconi (2014), composed of 13 texture classes (2080 images

with different rotations) taken from the Brodatz’s album (Brodatz,

1966). The image with rotation 0° are used in the training set, and

the rotated images belong to the testing set.• PI: the dataset in Borghesani, Grana, and Cucchiara (2014), con-

taining pictures extracted from digitalized pages of the Holy Bible

of Borso d’Este, duke of Ferrara (Italy) from 1450 A.D. to 1471 A.D.

The dataset PI is composed of 13 classes that are characterized by

clear semantic meaning and significant search relevance.• PA: the dataset in Khan, Beigpour, Weijer, and Felsberg (2014),

containing 2338 paintings by 50 painters, representative of 13 dif-

ferent painting styles: abstract expressionism, baroque, construc-

tivism, cubism, impressionism, neo-classical, pop art, post im-

pressionism, realism, renaissance, romanticism, surrealism, and

symbolism. All experiments on PA use the same training and test-

ing sets that were used in Khan et al. (2014) (these sets were pro-

vided by the authors).

An important goal in our research efforts is to produce general-

urpose systems that perform well across very different image

atasets and problems. For this reason we test our algorithms on

atasets, such as those listed above, that are intentionally very gen-

ral and different from each other. The presence of too many datasets

f the same kind would have polarized the overall results. Moreover,

or a system to be truly general-purpose, it is essential that all param-

ters and weights remain constant across all datasets in our experi-

ental section. This is crucial in order to avoid overfitting on single

atasets.

The weights of the different weighted sum rules proposed in

his paper are obtained using a grid search (where each weight

etween 0 and 5 is tested in increments of 0.5) for maxi-

izing the performance in the three datasets used in Nanni

t al. (2013b); Nanni et al. (2013c). Also the parameters of the prepro-

essing methods are obtained in a similar way. The results on these

atasets are not reported in this paper since these datasets were used

or parameter selection only. The three datasets used to determine

he weighted sum rules are the Smoke datasets (Feiniu, 2011), the

ocate Transfected dataset (Hamilton, Pantelic, Hanson, & Teasdale,

007), and RNAi, a dataset of fluorescence microscopy images of fly

ells (Zhang & Pham, 2011).

. Experimental results

We used the 5-fold cross-validation protocol to test each texture

escriptor, except for the VIR, PR, and PA datasets, which require the

L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000 8997

Table 2

Usefulness of the preprocessing methods.

HR WAV LBP GRA OR WLD LPQ DJ F1 F2 F3 F4

PAP 89.2 90.3 84.3 87.3 87.8 90.2 84.6 89.1 90.6 90.1 89.3

VIR 96.6 96.7 93.1 96.0 96.8 96.4 89.5 97.8 98.0 98.1 98.1

CHO 99.8 99.9 99.3 99.8 99.9 99.1 99.8 99.9 99.9 99.9 99.9

HI 89.6 90.2 85.1 88.9 89.8 87.3 86.6 90.3 90.6 90.8 90.9

BC 94.2 89.9 90.3 95.4 93.3 91.9 80.4 95.3 94.5 95.4 94.6

PR 89.9 82.1 87.9 88.6 84.0 79.8 85.3 90.1 87.8 88.5 89.8

HE 97.4 98.2 97.2 97.8 96.4 96.4 97.3 97.6 97.8 98.1 98.3

LO 99.5 99.3 98.6 99.2 98.0 98.2 98.9 99.4 99.4 99.4 99.5

BR 92.1 86.9 86.7 91.7 87.5 89.3 94.4 90.1 90.9 91.5 93.3

PI 87.9 81.1 85.6 87.8 85.5 86.7 87.1 88.0 88.1 88.1 88.1

PA 83.8 78.5 83.2 84.1 83.6 83.7 83.4 85.0 85.4 85.6 85.6

Av 92.7 90.2 90.1 92.4 91.1 90.8 89.7 92.9 93.0 93.2 93.4

SH WAV LBP GRA OR WLD LPQ DJ F1 F2 F3 F4

PAP 85.2 84.2 83.0 83.9 88.9 86.7 85.9 88.6 89.2 88.5 89.3

VIR 90.6 92.0 86.9 87.6 93.3 92.0 59.8 95.5 95.6 95.7 95.3

CHO 99.5 99.5 99.6 99.6 97.6 98.1 99.5 99.7 99.6 99.9 99.9

HI 84.2 85.2 83.1 85.6 87.8 76.9 81.9 89.4 89.3 90.2 90.7

BC 90.7 89.4 82.8 83.8 95.1 94.5 78.1 95.3 96.0 95.8 94.5

PR 89.3 82.0 85.7 90.2 87.2 79.0 85.0 91.3 90.2 91.6 91.7

HE 96.6 94.7 96.4 97.1 93.1 91.6 96.0 96.8 96.7 97.7 98.0

LO 98.8 88.4 98.9 98.4 96.8 95.0 93.1 99.1 99.0 99.3 99.2

BR 91.7 88.1 93.7 89.0 93.6 85.6 94.9 93.8 93.2 92.7 94.3

PI 80.5 74.5 80.2 81.0 81.1 76.5 76.5 82.7 82.8 83.1 83.0

PA 78.6 73.2 77.5 78.1 78.0 75.3 74.5 79.1 79.6 80.0 80.1

Av 89.6 86.4 87.9 88.5 90.2 86.4 84.1 91.9 91.9 92.2 92.3

GR WAV LBP GRA OR WLD LPQ DJ F1 F2 F3 F4

PAP 87.6 89.6 73.9 81.2 86.0 91.0 87.0 88.4 90.3 89.9 90.8

VIR 94.3 95.0 88.6 93.8 94.9 95.9 86.1 96.6 97.2 97.4 97.3

CHO 97.9 99.9 96.9 98.1 99.9 99.9 98.8 99.9 99.9 99.9 99.9

HI 86.9 90.4 81.0 83.9 87.6 86.9 85.0 89.8 91.4 91.5 91.8

BC 95.3 96.8 86.8 94.8 97.3 96.7 77.7 97.8 98.1 98.3 97.4

PR 89.9 91.7 83.4 87.6 90.1 89.8 87.1 92.6 93.3 93.0 93.5

HE 96.4 97.8 95.7 97.2 97.4 96.4 97.7 97.9 98.2 98.5 98.8

LO 98.8 99.6 96.5 98.0 99.1 98.8 98.7 99.6 99.6 99.6 99.6

BR 90.4 94.2 90.6 93.2 93.8 90.3 91.4 93.4 94.4 94.9 94.8

PI 80.5 81.4 79.5 81.1 80.7 78.4 78.5 92.8 93.0 83.2 83.2

PA 79.2 78.2 78.0 79.8 78.5 77.5 77.1 81.5 81.9 82.1 82.0

Av 90.6 92.2 86.4 89.8 91.3 91.0 87.7 93.6 94.3 93.4 93.5

SD WAV LBP GRA OR WLD LPQ DJ F1 F2 F3 F4

PAP 87.9 88.9 80.4 84.9 88.5 89.4 81.0 88.7 89.8 89.6 89.3

VIR 92.0 93.5 83.4 89.5 94.0 93.4 75.8 95.6 96.1 96.0 95.9

CHO 99.6 99.8 98.6 99.8 99.7 99.5 95.0 99.9 99.9 99.9 99.9

HI 85.6 87.0 80.4 84.9 86.8 84.3 79.6 88.6 89.8 90.4 90.4

BC 95.0 94.4 90.0 91.0 94.0 96.1 76.4 96.4 96.9 96.8 95.8

PR 90.3 87.2 88.2 90.7 84.8 85.7 77.6 90.5 90.5 91.3 90.7

HE 97.3 96.7 94.8 97.3 96.1 96.0 96.5 97.8 97.9 98.5 98.5

LO 99.3 99.1 98.3 99.4 98.2 97.4 97.1 99.3 99.3 99.5 99.5

BR 91.3 93.4 88.9 90.1 86.9 90.4 93.4 90.3 91.7 92.1 93.6

PI 80.0 81.5 81.2 81.0 79.5 78.6 77.2 82.2 82.8 83.2 83.2

PA 79.2 78.6 78.4 79.7 77.2 79.1 76.5 81.5 82.0 82.1 82.0

Av 90.6 90.9 87.5 89.8 89.6 89.9 84.1 91.8 92.4 92.6 92.6

o

u

p

m

w

t

c

(

a

m

p

e

t

p

t

S

i

fi

w

riginal testing protocols that were detailed in Section 4. The area

nder the ROC curve (AUC)5 is the performance indication since it

rovides a better overview of classification results. AUC is a scalar

easure that can be interpreted as the probability that the classifier

ill assign a higher score to a randomly picked positive sample than

o a randomly picked negative sample (Fawcett, 2004). In the multi-

lass problem, AUC is calculated using the one-versus-all approach

i.e., a given class is considered “positive” while all the other classes

re considered “negative”) and the average AUC is reported.

The aim of the first experiment, reported in Table 2, is to deter-

ine the usefulness of the different preprocessing techniques ap-

lied to the image before creating the co-occurrence matrix. In this

xperiment all three feature extraction approaches (see Section 2) are

ested. In addition, the following fusions are reported:

5 AUC is implemented as in dd_tools 0.95 davidt@ph.tn.tudelft.nl.

a

t

d

F1: WAV + WLD;

F2: WAV + WLD + LPQ;

F3: WAV + WLD+LPQ + OR;

F4: WAV + WLD+LPQ + OR + DJ.

A method represented above as A + B is the sum rule between ap-

roaches A and B. Before fusion, the scores of A and B are normalized

o mean 0 and standard deviation 1.

For feature extractors that we tested are HRsca, SHsca, GRsca and

Dsca. When a given preprocessing method produces more than one

mage (e.g., wavelet produces four different images), a SVM classi-

er is applied separately to each processed image. For instance, with

avelet we train one SVM using the approximation coefficient matrix

nd three SVMs using the three detail coefficient matrices (horizon-

al, vertical, and diagonal). This set of SVMs is combined by sum rule.

Notice that we report only the best ensemble (average in the four

escriptors) built by 2, 3, 4, and 5 preprocessing approaches. We want

8998 L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000

Table 3

Performance gains using the proposed ensemble.

AUC HR HRsca HRp HRsca+ 2×HRp SH SHsca SHp SHsca+ 2×SHp GR GRsca GRp GRsca+ 2×GRp SD SD sca SDp SDsca+ 2× SDp

PAP 89.4 92.5 90.1 92.4 82.5 86.5 88.5 89.9 84.1 85.9 89.9 89.9 84.8 86.7 89.6 89.6

VIR 95.9 96.7 98.1 98.4 84.6 92.1 95.7 96.7 88.8 94.1 97.4 97.7 76.8 88.4 96.0 96.4

CHO 99.4 99.5 99.9 99.9 99.5 99.8 99.9 99.9 98.6 97.9 99.9 99.9 98.7 99.4 99.9 99.9

HI 87.7 89.5 90.8 91.6 83.2 89.6 90.2 91.7 87.3 89.6 91.5 92.8 81.2 87.1 90.4 90.7

BC 92.7 93.8 95.4 95.6 88.8 92.3 95.8 95.8 84.9 93.8 98.3 98.0 89.4 91.5 96.8 95.9

PR 90.6 91.0 88.5 91.0 84.8 87.6 91.6 91.0 85.7 92.7 93.0 94.2 79.2 85.7 91.3 90.6

HE 96.1 97.2 98.0 98.1 94.1 96.7 97.7 98.0 92.2 97.0 98.5 98.7 95.3 97.7 98.5 98.8

LO 98.1 99.1 99.2 99.5 95.4 98.1 99.3 99.3 98.4 99.0 99.6 99.6 97.6 98.8 99.5 99.6

BR 88.9 91.3 93.3 91.8 92.7 96.6 92.7 94.9 89.9 93.1 94.9 95.3 88.8 93.2 92.1 93.0

PI 87.7 90.4 88.1 88.3 81.0 85.0 83.1 85.1 81.1 85.7 83.2 85.9 81.0 85.1 83.2 86.0

PA 84.1 87.5 85.6 87.6 78.1 83.4 80.0 84.0 79.8 84.1 82.1 84.8 79.7 83.9 82.1 84.8

Av 91.8 93.5 93.3 94.0 87.7 91.6 92.2 93.3 88.2 92.0 93.4 94.2 86.5 90.6 92.6 93.2

Table 4

Performance of fusion approaches.

AUC HR NanniA NanniB NanniC SUMdiv SUM WEI

PAP 89.4 89.5 92.5 91.8 91.6 91.7 91.8

VIR 95.9 96.0 96.4 96.5 97.9 98.1 98.2

CHO 99.4 99.5 99.9 99.9 99.9 100 99.9

HI 87.7 88.0 90.7 92.4 93.5 93.5 93.6

BC 92.7 92.8 94.9 95.1 97.2 97.1 97.1

PR 90.6 90.6 89.1 92.1 93.0 92.8 93.0

HE 96.1 96.5 97.8 98.3 98.8 98.6 98.7

LO 98.1 98.8 99.4 99.5 99.6 99.6 99.6

BR 88.9 90.0 93.3 94.9 94.1 94.7 94.6

PA 84.1 85.1 85.5 86.2 87.8 88.0 88.1

PI 87.8 88.2 88.8 89.2 90.2 90.4 90.8

Av 91.9 92.3 93.5 94.2 94.9 94.9 95.0

Table 5

Fusion approaches.

AUC LBP LTP LPQ CLBP MT W1 ENS

PAP 87.7 86.1 90.2 92.5 90.0 92.6 92.0

VIR 89.8 91.6 94.9 94.8 93.7 96.5 98.2

CHO 99.9 100 99.2 99.9 100 100 100

HI 92.5 92.8 92.0 91.8 93.4 94.1 94.3

BC 92.4 95.6 95.7 95.8 96.0 97.3 97.7

PR 79.8 87.8 86.2 86.6 88.4 93.2 93.6

HE 98.0 98.6 97.2 98.1 98.3 98.7 98.8

LO 98.6 99.5 97.6 98.5 99.6 99.6 99.7

BR 86.5 90.9 91.7 90.2 95.6 96.7 96.5

PA 87.3 89.0 88.3 93.2 90.1 90.3 90.8

PI 90.1 92.9 90.7 88.6 93.2 93.5 93.9

Av 91.1 93.2 93.1 93.6 94.4 95.7 96.0

i

f

i

e

i

l

S

p

f

6 For LBP, LTP, CLBP, and MT, we use uniform bins with (P = 8, R = 1) and (P = 16,

R = 2); for LPQ the size of the local window was {3, 5}.

to stress that it is well known in the literature that a good ensemble

is obtained by considering a tradeoff between the performance and

the diversity of the individual classifiers that make up the ensemble

(Kuncheva, 2014). For this reason, LBP is not used in any ensemble

due to its lower diversity with respect to WLD and LPQ.

Examining Table 2, we see that the performance of F3 and F4 are

similar. In F3, the DJ preprocessing method, which is computationally

expensive, is not used. In the tests that follow, therefore, we use F3 to

reduce computation time.

The goal of the second experiment, reported in Table 3, is to show

the performance improvements obtainable when the features ex-

tracted from the original image are combined with the preprocessed

images. The following methods are compared in Table 3:

• {HR, SH, GR, SD} is the performance of the descriptors {HR, SH, GR,

SD} extracted from the whole original image only;• {HRsca, SHsca, GRsca, SDsca} is the performance of the descriptors

{HRsca, SHsca, GRsca, SDsca};• {HRp, SHp, GRp, SDp} is the best fusion reported

(WAV+WLD+LPQ+OR) in Table 2 by varying the types of fea-

tures extracted from the processed images;• A + K × B is the fusion by weighted sum rule between A and B

(weight of B is K). Before fusion, the scores of A and B are normal-

ized to mean 0 and standard deviation 1.

Examining Table 3, we see that the weighted sum rule outper-

forms all the other approaches. It is interesting to note the improve-

ment obtained with respect to the baseline approaches, i.e., where

the features are extracted from the whole co-occurrence matrix us-

ing the original image only (i.e., HR, SH, GR and SD).

To statistically validate our experiments, we used the Wilcoxon

signed rank test (Demšar, 2006), which confirms (p-value 0.05) that

the proposed fusion outperforms the standard approaches HR, SH,

GR, and SD.

The aim of the third experiment, reported in Table 4, is to exam-

ne the performance gained by fusing different descriptors extracted

rom the co-occurrence matrix. The methods compared in this exper-

ment are the following:

• NanniA: the best ensemble proposed in Nanni et al. (2013b);• NanniB: the best ensemble proposed in Nanni et al. (2013c);• NanniC: the best ensemble proposed in Nanni et al. (2013d);• SUMdiv: proposed in this paper as the sum rule fusion of

HRsca + 2 × HRp, GRsca + 2 × GRp and DIsca + 2 × DIp.• SUM: proposed in this paper as the sum rule fusion of

HRsca + 2 × HRp, GRsca + 2 × GRp, and SHsca + 2 × SHp.• WEI: proposed in this paper as the sum rule fusion of

HRsca + 2 × HRp (weight 2), GRsca + 2 × GRp (weight 2), and

SHsca + 2 × SHp (weight 1).

Examining Table 4, we find WEI outperforming all our previous

nsembles, NanniA, NanniB and NanniC, with a p-value of 0.05 us-

ng Wilcoxon signed rank test. SUMdiv obtains a performance simi-

ar to SUM. In our experiments, we tried to improve WEI by adding

Dsca + 2 × SDp to the ensemble but obtained no improvement in

erformance.

Finally, in Table 5, we report the results of some of the best per-

orming texture descriptors reported in the literature:

• Local binary patterns (LBP)6 (Ojala et al., 2002).• Local ternary patterns (LTP) (Tan & Triggs, 2007).• Multi-threshold local quinary coding (MT) (Paci et al., 2013).• Local phase quantization (LPQ) (Rahtu et al., 2012).• Completed Local binary patterns (CLBP) (Guo, Zhang, & Zhang,

2010).• W1: an ensemble of descriptors proposed in Nanni et al. (2013d).

L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000 8999

f

6

m

i

W

t

f

T

s

i

s

c

c

t

w

p

H

f

b

s

p

t

a

t

p

t

p

c

f

p

s

s

p

t

w

G

m

L

R

A

B

B

B

B

B

C

C

C

D

D

D

D

F

F

F

F

G

G

G

G

G

H

H

H

H

H

J

J

J

K

K

K

L

M

M

M

N

N

N

N

Moreover we report the following fusion:

• ENS: sum rule between MT and WEI.

Examining Table 5, we see that the proposed system ENS outper-

orms the other approaches.

. Conclusion

In this study we improved the performance of the co-occurrence

atrix by examining different preprocessing techniques applied to

mages before extracting features from the co-occurrence matrix.

e also compared and combined strategies for extending the tex-

ure descriptors extracted from the co-occurrence matrix after trans-

orming the original image with different preprocessing methods.

hese methods are then improved by combining them with a multi-

cale approach that is based on Gaussian filtering and on extract-

ng features from both the entire co-occurrence matrix and a set of

ub-windows. For all experiments, the SVM was used as the base

lassifier.

Our proposed system was validated across several image classifi-

ation problems, with very different images, thereby demonstrating

he generalizability of our approach. We also compared our results

ith some state-of-the-art descriptors. In our opinion, the results re-

orted in this paper are very useful for practitioners; rather than use

aralick’s descriptors, they can now use our more powerful set of

eatures extracted from the GLCM. The code of the proposed ensem-

le is available at https://www.dei.unipd.it/node/2357. Moreover, our

tudy shows the value of exploring novel methods for deriving more

owerful descriptors from the co-occurrence matrix.

Because one of our goals was to create a general-purpose system

hat is capable of working with many image problems and databases,

limitation of this work could be the number of tested datasets. In

he future we plan on expanding tests so that our systems are com-

ared on at least 15 datasets representing different image classifica-

ion problems. This expansion would provide a more reliable com-

arison. Another limitation of our proposed system has to do with

omputation time: the standard approach based on the Haralick’s

eatures is considerably less expensive, computationally, than our

roposed system since it requires the training of several SVMs. It

hould be pointed out, however, that each SVM in our proposed en-

emble is trained independently from the others, so the proposed ap-

roach is well suited for modern multi-core architectures.

As a follow-up to this work, we plan on exploring recent methods

hat boost the performance of texture descriptors, e.g., approaches

here each image is split into different subimages (Vu, Nguyen, &

arcia, 2014). We will also examine new methods for choosing the

ost powerful set of state-of-the-art texture descriptors (such as,

TP+LQP) that can be extracted from the co-occurrence matrix.

eferences

kono, A., Tankam, N. T., & Tonye, E. (2005). Nouvel algorithme d’evaluation de texture

d’ordre n sur la classification de l’ocupation des sols de la region volcanique dumont Cameroun. Teledetection, 5, 227–244.

ay, H., Tuytelaars, T., & Gool, L. V. (2006). SURF: speeded up robust features. EuropeanConference on Computer Vision, 1, 404–417.

oland, M. V., Markey, M. K., & Murphy, R. F. (1998). Automated recognition of pat-

terns characteristic of subcellular structures in fluorescence microscopy images.Cytopathology, 33, 366–375.

oland, M. V., & Murphy, R. F. (2001). A neural network classifier capable of recognizingthe patterns of all major subcellular structures in fluorescence microscope images

of HeLa cells. Bioinformatics, 17(12), 1213–1223.orghesani, D., Grana, C., & Cucchiara, R. (2014). Miniature illustrations retrieval and

innovative interaction for digital illuminated manuscripts in multimedia systems.Multimedia Systems, 20, 65–79.

rodatz, P. (1966). Textures: a photographic album for artists and designers. New York,

USA: Dover Publications.hen, S., Chengdong, W., Chen, D., & Tan, W. (2009). Scene classification based on gray

level-gradient co-occurrence matrix in the neighborhood of interest points. In Pro-ceedings of the ieee international conference on intelligent computing and intelligent

systems 2009 (pp. 482–485).

hen, J., Shan, S., He, C., Zhao, G., & Pietikäinen, M. (2010). WLD: a robust local im-age descriptor. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(9),

1705–1720.ruz-Roa, A., Caicedo, J. C., & González, F. A. (2011). Visual pattern mining in histology

image collections using bag of features. Artificial Intelligence in Medicine, 52(2), 91–106.

alal, N., & Triggs, B. (2005). Histograms of oriented gradients for human detection. InProceedings of the 9th European Conference on Computer Vision (pp. 886–893).

e Mesquita Sá Junior, J. J., Backes, A. R., & Cortez, P. C. (2013). Texture analysis and

classification using shortest paths in graphs. Pattern Recognition Letters, 34, 1314–1319.

emšar, J. (2006). Statistical comparisons of classifiers over multiple data sets. Journalof Machine Learning Research, 7, 1–30.

ijkstra, E. W. (1959). A note on two problems in connexion with graph. NumerischeMathematik, 1, 269–271.

awcett, T. (2004). ROC graphs: notes and practical considerations for researchers. Palo

Alto, USA: HP Laboratories.einiu, Y. (2011). Video-based smoke detection with histogram sequence of LBP and

LBPV pyramids. Fire Safety Journal, 46(3), 132–139.ink, J. L., Aturaliya, R. N., Davis, M. J., Zhang, F., Hanson, K., Teasdale, M. S., et al. (2006).

LOCATE: a mouse protein subcellular localization database. Nucleic Acids Research,34, 213–217.

reeman, W. T., & Adelson, E. H. (1991). The design and use of steerable filters. IEEE

Transactions on Pattern Analysis and Machine Intelligence, 13, 891–906.elzinis, A., Verikas, A., & Bacauskiene, M. (2007). Increasing the discrimination

power of the co-occurrence matrix-based features. Pattern Recognition, 40, 2367–2372.

eraldo Junior, B., da Rocha, S. V., Gattass, M., Silva, A. C., & de Paiva, A. C. (2013). A massclassification using spatial diversity approaches in mammography images for false

positive reduction. Expert System with Applications, 40, 7534–7543.

hidoni, S., Cielniak, G., & Menegatti, E. (2012). Texture-based crowd detection andlocalisation. In International Conference on Intelligent Autonomous Systems 2012: 193

(pp. 725–736).onzalez, E., Fernandez, A., & Bianconi, F. (2014). General framework for rotation in-

variant texture classification through co-occurrence of patterns. Journal of Mathe-matical Imaging and Vision, 50(3), 300–313.

uo, Z., Zhang, L., & Zhang, D. (2010). A completed modeling of local binary pattern op-

erator for texture classification. IEEE Trans. on Image Processing, 19(6), 1657–1663.amilton, N., Pantelic, R., Hanson, K., & Teasdale, R. D. (2007). Fast automated cell phe-

notype classification. BMC Bioinformatics, 8–110.aralick, R. M., Shanmugam, K., & Dinstein, I. (1973). Textural features for image clas-

sification. IEEE Transactions on Systems, Man, and Cybernetics, 3, 610–621.aralick, R. M. (1979). Statistical and structural approaches to texture. In Proceedings of

the IEEE: 67 (pp. 786–804).

u, Y. (2009). Unsupervised texture classification by combining multi-scale featuresand k-means classifier. In Proceedings of the Chinese Conference on Pattern Recogni-

tion (pp. 1–5).uang, J., Kumar, S. R., Mitra, M., Zhu, W. J., & Zabih, R. (1997). Image indexing using

color correlograms. In Proceedings of the IEEE Conference on Computer Vision andPattern Recognition (pp. 762–768).

ain, A. K., & Farrokhnia, F. (1991). Unsupervised texture segmentation using gabor fil-ters. Pattern Recognition Letters, 24, 1167–1186.

antzen, J., Norup, J., Dounias, G., & Bjerregaard, B. (2005). Pap-smear benchmark data

for pattern classification. Nature inspired Smart Information Systems, 1, 1–9.unior, G. B., Cardoso de Paiva, A., Silva, A. C., & Muniz de Oliveira, A. C. (2009). Clas-

sification of breast tissues using Moran’s index and Geary’s coefficient as texturesignatures and SVM. Computers in Biology and Medicine, 39, 1063–1072.

han, F., Beigpour, S., Weijer, J., & Felsberg, M. (2014). Painting-91: a large scaledatabase for computational painting categorization. Machine Vision and Applica-

tions, 25, 1385–1397.

uncheva, I. L. (2014). Combining pattern classifiers: methods and algorithms (2nd Edi-tion). New York: Wiley, ISBN: 987-1-118-31523-1, 384 pages, October 2014.

ylberg, G., Uppström, M., & Sintorn, I.-M. (2011). Virus texture analysis using localbinary patterns and radial density profiles. In Proceedings of the 18th Iberoamerican

Congress on Pattern Recognition (pp. 573–580).owe, D. G. (2004). Distinctive image features from scale-invariant keypoints. Interna-

tional Journal of Computer Vision, 60, 91–110.

allat, S. (1989). A theory for multiresolution signal decomposition. IEEE Transactionson Pattern Analysis and Machine Intelligence, 11, 674–693.

ikolajczyk, K., & Schmid, C. (2005). A performance evaluation of local descriptors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 29, 1615–1630.

itrea, D., Mitrea, P., Nedevschi, S., Badea, R., Lupsor, M., et al. (2012). Abdominal tu-mor characterization and recognition using superior-order cooccurrence matrices,

based on ultrasound images. Computational and Mathematical Methods in Medicine,

2012, 1–17.anni, L., & Lumini, A. (2008). A reliable method for cell phenotype image classification.

Artificial Intelligence in Medicine, 43, 87–97.anni, L., Shi, J.-Y., Brahnam, S., & Lumini, A. (2010). Protein classification using tex-

ture descriptors extracted from the protein backbone image. Journal of TheoreticalBiology, 3, 1024–1032.

anni, L., Paci, M., Brahnam, S., Ghidoni, S., & Menegatti, E. (2013a). Virus image classi-

fication using different texture descriptors. In The 14th International Conference onBioinformatics and Computational Biology (pp. 56–61).

anni, L., Brahnam, S., Ghidoni, S., Menegatti, E., & Barrier, T. (2013b). A comparison ofmethods for extracting information from the co-occurrence matrix for subcellular

classification. Expert Systems with Applications, 40(18), 7457–7467.

9000 L. Nanni et al. / Expert Systems With Applications 42 (2015) 8989–9000

S

S

S

T

T

T

U

V

V

V

V

V

W

W

Z

Z

Nanni, L., Ghidoni, S., & Menegatti, E. (2013c). A comparison of multi-scale approachesfor extracting image descriptors from the co-occurrence matrix. Computer Commu-

nication & Collaboration, 1(1), 1–18.Nanni, L., Brahnam, S., Ghidoni, S., & Menegatti, E. (2013d). Different approaches for

extracting information from the co-occurrence matrix. PLoS ONE, 8(12), e83554.doi:10.1371/journal.pone.0083554.

Ojala, T., Pietikäinen, M., & Harwood, D. (1996). A comparative study of texture mea-sures with classification based on featured distribution. Pattern Recognition Letters,

29, 51–59.

Ojala, T., Pietikainen, M., & Maeenpaa, T. (2002). Multiresolution gray-scale and rota-tion invariant texture classification with local binary patterns. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 24, 971–987.Ojansivu, V., & Heikkila, J. (2008). Blur insensitive texture classification using lo-

cal phase quantization. In International Conference on Image and Signal Processing(pp. 236–243).

Oliver, A., Lladó, X., Freixenet, J., & Martí, J. (2007). False positive reduction in mammo-

graphic mass detection using local binary patterns. Medical Image Computing andComputer-Assisted Intervention, 4791, 286–293.

Paci, M., Nanni, L., Lathi, A., Aalto-Setälä, K., & Hyttinen, J. (2013). Non-binary codingfor texture descriptors in sub-cellular and stem cell image classification. Current

Bioinformatics, 8, 208–219.Pacifici, F., & Chini, M. (2008). Urban land-use multi-scale textural analysis. In Proceed-

imgs of the ieee international geoscience and remote sensing symposium (pp. 342–

345).Pantic, I., Dacic, S., Brkic, P., Lavrnja, I., Pantic, S., & Jovanovic, T. (2014). Application

of fractal and grey level co-occurrence matrix analysis in evaluation of brain cor-pus callosum and cingulum architecture. Microscopy and Microanalysis: The Official

Journal of Microscopy Society of America, Microbeam Analysis Society, MicroscopicalSociety of Canada, 20(5), 1373–1381.

Pass, G., Zabih, R., & Miller, J. (1996). Comparing images using color coherence vectors.

ACM International Conference on Multimedia, 65–73.Rahtu, E., Heikkilä, J., Ojansivu, V., & Ahonen, T. (2012). Local phase quantization for

blur-insensitive image analysis. Image and Vision Computing, 30(8), 501–512.Rakwatin, P., Longepe, N., Isoguchi, O., Shimada, M., & Uryu, Y. (2010). Mapping tropical

forest using ALOS PALSAR 50 m resolution data with multiscale GLCM analysis.In Proceedings of the ieee international geoscience and remote sensing symposium

(pp. 1234–1237).

Ren, X.-M., Wang, X.-F., & Zhao, Y. (2012). An efficient multi-scale overlapped blockLBP approach for leaf image recognition. In Proceedings of the intelligent computing

theories and applications (pp. 237–243).Sastry, S. S., Kumari, T. V., Rao, C. N., Mallika, K., & Lakshminarayana, S. (2012). Transi-

tion temperatures of thermotropic liquid crystals from the local binary gray levelco-occurrence matrix. Advances in Condensed Matter Physics, 2012, 1–9.

ubrahmanyam, M. J., Wu, Q. M., Maheshwari, R. P., & Balasubramanian, R. (2013). Mod-ified color motif co-occurrence matrix for image indexing and retrieval. Computers

& Electrical Engineering, 39(3), 762–774.un, X., Wang, J., Chen, R., She, M. F. H., & Kong, L. (2012). Multi-scale local pattern co-

occurrence matrix for textural image classification. In Proceedings of the ieee worldcongress on computational intelligence (pp. 1–7).

wain, M. J., & Ballard, D. H. (1991). Color indexing. International Journal of ComputerVision, 7, 11–32.

an, X., & Triggs, B. (2007). Enhanced local texture feature sets for face recognition un-

der difficult lighting conditions. Analysis and Modelling of Faces and Gestures LNCS,4778, 168–182.

ang, X. (1998). Texture information in run-length matrices. IEEE Transactions On ImageProcessing, 7(11), 1602–1609.

uzel, O., Porikli, F., & Meer, P. (2006). Region covariance: a fast descriptor for detectionand classification. In Proceedings of the 9th european conference on computer vision

(pp. 589–600).

nay, D., & Ekin, A. (2008). Intensity versus texture for medical image search and re-trieval. In Proceedings of the 5th ieee international symposium on biomedical imaging:

from nano to macro (pp. 241–244).apnik, V., & Lerner, A. (1963). Pattern recognition using generalized portrait method.

Automation and Remote Control, 24, 774–780.apnik, V., & Chervonenkis, A. (1964). A note on one class of perceptrons. Automation

and Remote Control, 25.

écsei, A., Amann, G., Hegenbart, S., Liedlgruber, M., & Uhl, A. (2011). Automatedmarsh-like classification of celiac disease in children using local texture operators.

Computers in Biology and Medicine, 41, 313–325.u, N.-S. (2013). Exploring patterns of gradient orientations and magnitudes for

face recognition. IEEE Transactions on Information Forensics and Security, 8, 295–304.

u, N., Nguyen, T., & Garcia, C. (2014). Improving texture categorization with

biologically-inspired filtering. Image and Vision Computing, 32(6–7), 424–436.

alker, R., Jackway, P., & Longstaff, D. (2003). Genetic algorithm optimization of adap-tive multi-scale GLCM features. International Journal of Pattern Recognition and Ar-

tificial Intelligence, 17, 17–39.u, B., & Nevatia, R. (2005). Detection of multiple, partially occluded humans in a sin-

gle image by bayesian combination of edgelet part detectors. In Proceedings of the

IEEE International Conference on Computer Vision (pp. 90–97).hang, B., & Pham, T. D. (2011). Phenotype recognition with combined fea-

tures and random subspace classifier ensemble. BMC Bioinformatics, 12(1),128.

hao, L., Peng, T., Han, H., Cao, W., Lou, Y., & Xie, X. (2014). Fault condition recognitionbased on multi-scale co-occurrence matrix for copper flotation process. In Proceed-

ings of the 19th world congress of the international federation of automatic control

(pp. 7091–7097).