Local Texton Dissimilarity with Applications on Biomass Classiﬁcation

Radu Tudor Ionescu

, Andreea-Lavinia Popescu

, Dan Popescu

and Marius Popescu

Faculty of Mathematics and Computer Science, University of Bucharest,

14 Academiei Street, Bucharest, Romania

Faculty of Automatic Control and Computer Science, Politehnica University of Bucharest,

313 Splaiul Independentei Street, Bucharest, Romania

Keywords:

Texture Dissimilarity, Texture Classiﬁcation, Biomass Classiﬁcation, Biomass Type Identiﬁcation, Textons,

Texton-based Technique.

Abstract:

Texture classiﬁcation, texture synthesis, or similar tasks are an active topic in computer vision and pattern

recognition. This paper aims to present a novel texture dissimilarity measure based on textons, namely the

Local Texton Dissimilarity (LTD), inspired from (Dinu et al., 2012). Textons are represented as a set of fea-

tures extracted from image patches. The proposed dissimilarity measure shows its application on biomass

type identiﬁcation. A new data set of biomass texture images is provided by this work, which is available

at http://biomass.herokuapp.com. Images are separated into three classes, each one representing a type of

biomass. The biomass type identiﬁcation and quality assessment is of great importance when one in the

biomass industry needs to produce another energy product, such as biofuel, for example. Two more experi-

ments are conducted on popular texture classiﬁcation data sets, namely Brodatz and UIUCTex. The proposed

method beneﬁts from a faster computational time compared to (Dinu et al., 2012) and a better accuracy when

used for texture classiﬁcation. The performance level of the machine learning methods based on LTD is

comparable to the state of the art methods.

1 INTRODUCTION

Computer vision researchers have developed sophis-

ticated methods for image related tasks, such as im-

age retrieval, image categorization, image segmenta-

tion or image synthesis. Indeed, many of the most

powerful of these methods are patch-based (Barnes

et al., 2011). Such methods divide the image into

many small patches and then manipulate or analyze

the image based on its patches. Some of these meth-

ods have been designed or adapted to work on tex-

tures, where vector quantized image patches are also

referred to as textons (Leung and Malik, 2001), (Xie

et al., 2010).

Texture classiﬁcation, texture synthesis, or similar

tasks are an active topic in computer vision and pat-

tern recognition, having many practical applications.

This paper aims to present a novel texture dissimilar-

ity measure based on textons, namely the Local Tex-

ton Dissimilarity (LTD), inspired from (Dinu et al.,

2012). Textons are represented as a set of features ex-

tracted from image patches. Similar textons will be

represented through similar features. Thus, images

patches are implicitly quantized into textons. Textons

provide a lighter representation of patches, allowing

for a faster computational time and broader applica-

tion to practical problems. LTD sums the spatial off-

sets of similar textons to measure the similarity be-

tween two texture images.

Several experiments are conducted on three tex-

ture data sets. LTD shows its ﬁrst application on

biomass type identiﬁcation, a direct application of

texture classiﬁcation. A new data set of biomass tex-

ture images is provided by this work. Images are sep-

arated into three classes, each one representing a type

biomass. A method to determine the biomass type

has practical motivations for the biomass industry.

Such methods are of great importance when one in

the biomass industry needs to produce another energy

product, such as biofuel or bioenergy, for example. Is

the type of biomass appropriate to efﬁciently obtain

the bioproduct? Is the biomass conversion method

the right one for this type of biomass? Answering

such questions can help reduce the operating costs of

biomass power plants. But, such questions can be an-

swered with the help of a simple biomass type iden-

tiﬁcation method, such as the one presented in this

work. The other experiments are conducted on two

593

Ionescu R., Popescu A., Popescu D. and Popescu M..

Local Texton Dissimilarity with Applications on Biomass Classiﬁcation.

DOI: 10.5220/0004740105930600

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 593-600

ISBN: 978-989-758-003-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

popular texture classiﬁcation data sets, namely Bro-

datz and UIUCTex. The proposed method beneﬁts

from a faster computational time compared to (Dinu

et al., 2012) and a better accuracy when used for tex-

ture classiﬁcation. The performance level of the ma-

chine learning methods based on LTD is comparable

to the state of the art methods.

The paper is organized as follows. Related work

about texton-based and patch-based techniques, and

biomass classiﬁcation is discussed in Section 2. LTD

is described in Section 3. Experiments with machine

learning methods based on LTD are presented in Sec-

tion 4. Finally, the conclusions are drawn in Section 5.

2 RELATED WORK

2.1 Patches and Textons

For numerous computer vision applications, the im-

age can be analyzed at the patch (or texton) level

rather than at the individual pixel level or global level.

Patches and textons contain contextual information

and have advantages in terms of computation and gen-

eralization. For example, patch-based methods pro-

duce better results and are much faster than pixel-

based methods for texture synthesis (Efros and Free-

man, 2001). However, patch-based techniques are

still heavy to compute with current machines (Barnes

et al., 2011).

The authors of (Lazebnik et al., 2005b) develop

a texture representation that is invariant to geomet-

ric transformations based on descriptors deﬁned on

afﬁne invariant regions. A probabilistic part-based ap-

proach for texture and object recognition is presented

in (Lazebnik et al., 2005a). Textures are represented

using a part dictionary obtained by quantizing the ap-

pearance of salient image regions.

In (Leung and Malik, 2001) texture images are

classiﬁed by using 3D textons, which are cluster cen-

ters of ﬁlter response vectors corresponding to differ-

ent lighting and viewing directions of images. The au-

thors of (Varma and Zisserman, 2005) model textures

by the joint distribution of ﬁlter responses. This dis-

tribution is represented by the frequency histogram of

textons. For most texton based techniques, the textons

are usually learned by k-means clustering. In (Xie

et al., 2010) the authors propose a novel texture clas-

siﬁcation method via patch-based sparse texton learn-

ing. The dictionary of textons is learned by applying

sparse representation to image patches in the training

data set. In (Barnes et al., 2011) the authors present

a new randomized algorithm for quickly ﬁnding ap-

proximate nearest neighbor matches between image

patches.

2.2 Biomass Classiﬁcation

In a general sense, biomass refers to the biological

material from living, or recently living organisms. In

this work, the term biomass refers to a renewable en-

ergy source, that can be directly converted into an-

other type of energy product. The Biomass Tex-

ture data set provided by this work is a collection of

close-up photos of different samples of three types of

biomass: municipal solid waste, corn, and wheat. The

goal is to build a classiﬁer that is able to distinguish

between these three types of biomass. This is a totally

different approach and understanding of the biomass

classiﬁcation problem, compared to other researches.

Usually, biomass classiﬁcation refers to land cover

type or forest biomass classiﬁcation. Land cover clas-

siﬁcation (Dash et al., 2007) and forest biomass esti-

mation (Wulder et al., 2008) are active research topics

in the area of remote sensing. The authors of (Hoek-

man and Quinnones, 2000) show that remotely sensed

image classiﬁcation systems may be designed to ac-

curately monitor processes of deforestation, land and

forest degradation and secondary forest regrowth.

3 LOCAL TEXTON

DISSIMILARITY

To compute LTD between two gray-scale texture im-

ages, the idea is to sum up all the offsets of similar

textons between the two images. The LTD algorithm

is brieﬂy described next. For every texton in one im-

age, the algorithm searches for a similar texton in the

other image. First, it looks for similar textons in the

same position in both textures. If those textons are

similar, it sums up 0 since there is no spatial offset

between textons. If the textons are not similar, the

algorithm starts exploring the vicinity of the initial

texton position in the second image to ﬁnd a texton

similar to the one in the ﬁrst image. If a similar tex-

ton is found during this process, it sums up the off-

set between the two textons. The spatial search goes

on until a similar texton is found or until a maximum

offset is reached. The maximum texton offset must be

set a priori. The computation of LTD is similar the al-

gorithm presented in (Dinu et al., 2012). In practice,

this computation is too heavy for a large set of im-

ages. To speed up the algorithm, textons are extracted

and compared using a dense grid over the image.

Notice that the algorithm proposed in (Dinu et al.,

2012) compares image patches by using the mean eu-

clidean distance. LTD differs in the way it compares

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

594

these patches. First, texture speciﬁc features are ex-

tracted from each patch. A patch can then be repre-

sented by a feature vector. Similar patches will be

represented through similar features. Thus, images

patches are implicitly quantized into textons. Textons

are compared using the Bhattacharyya coefﬁcient be-

tween their feature vectors. Section 3.1 describes the

features extracted from image patches. The algorithm

is presented in Section 3.2.

3.1 Texture Features

Before computing LTD between texture images, a set

of several image features is extracted from each patch

to obtain the texton representation. There are 9 fea-

tures extracted from patches, that are described next.

An interesting remark is that the more features are

added to the texton representation, the better the ac-

curacy of the LTD method gets. However, a lighter

representation, such as the one based on 9 features,

results in a faster and more efﬁcient algorithm. One

may choose to add or remove features in order to ob-

tain the desired trade-off between accuracy and speed.

The texton representation based on the 9 features that

are about to be presented next gives state of the art ac-

curacy levels in several experiments presented in Sec-

tion 4.

The ﬁrst two statistical features extracted are the

mean and the standard deviation. These two basic

features can be computed indirectly, in terms of the

image histogram. The shape of an image histogram

provides many clues to characterize the image, but the

features obtained from an image histogram are not al-

ways adequate to discriminate textures, since they are

unable to indicate local intensity differences.

One of the most powerful statistical methods for

textured image analysis is based on features extracted

from the Gray-Level Co-Occurrence Matrix (GLCM),

proposed in (Haralick et al., 1973). The GLCM is

a second order statistical measure of image variation

and it gives the joint probability of occurrence of gray

levels of two pixels, separated spatially by a ﬁxed vec-

tor distance. Smooth texture gives co-occurrence ma-

trix with high values along diagonals for small dis-

tances. The range of gray level values within a given

image determines the dimensions of a co-occurrence

matrix. Thus, 4 bits gray level images give 16 × 16

co-occurrence matrices. Relevant statistical features

for texture classiﬁcation can be computed from a

GLCM. The features proposed by (Haralick et al.,

1973), which show a good discriminatory power, are

the contrast, the energy, the entropy, the homogene-

ity, the variance and the correlation. Among these

features that show a good discriminatory power, LTD

uses only four of them, namely the contrast, the en-

ergy, the homogeneity, and the correlation.

Another feature that is relevant for texture analysis

is the fractal dimension. It provides a statistical index

of complexity comparing how detail in a fractal pat-

tern changes with the scale at which it is measured.

The fractal dimension is usually approximated. The

most popular method of approximation is box count-

ing (Falconer, 2003). The idea behind the box count-

ing dimension is to consider grids at different scale

factors over the fractal image, and count how many

boxes are ﬁlled over each grid. The box counting

dimension is computed by estimating how this num-

ber changes as the grid gets ﬁner by applying a box

counting algorithm. An efﬁcient box counting algo-

rithm for estimating the fractal dimension was pro-

posed in (Popescu et al., 2013). The idea of the algo-

rithm is to skip the computation for coarse grids, and

count how many boxes are ﬁlled only for ﬁner grids.

LTD includes this efﬁcient variant of box counting in

the texton representation.

The work of (Daugman, 1985) found that cells in

the visual cortex of mammalian brains can be mod-

eled by Gabor functions. Thus, image analysis by

the Gabor functions is similar to perception in the hu-

man visual system. A set of Gabor ﬁlters with differ-

ent frequencies and orientations may be helpful for

extracting useful features from an image. The lo-

cal isotropic phase symmetry measure (LIPSyM) pre-

sented in (Kuse et al., 2011) takes the discrete time

Fourier transform of the input image, and ﬁlters this

frequency information through a bank of Gabor ﬁl-

ters. The work of (Kuse et al., 2011) also notes that lo-

cal responses of each Gabor ﬁlter can be represented

in terms of energy and amplitude. Thus, Gabor fea-

tures, such as the mean-squared energy and the mean

amplitude, can be computed through the phase sym-

metry measure for a bank of Gabor ﬁlters with various

scales and rotations. These features are relevant be-

cause Gabor ﬁlters have been found to be particularly

appropriate for texture representation and discrimina-

tion.

Finally, textons are represented by the mean and

the standard deviation of the patch, the contrast, the

energy, the homogeneity, and the correlation extracted

from the GLCM, the (efﬁcient) box counting dimen-

sion, and the mean-squared energy and the mean am-

plitude extracted by using Gabor ﬁlters. These texton

features can be extracted from all images before com-

paring them with LTD. Thus, the LTD computation

can be divided in two main steps, one for texton fea-

ture extraction, and one for dissimilarity computation.

After the feature extraction step, features should be

normalized. In practice, the described features work

LocalTextonDissimilaritywithApplicationsonBiomassClassification

595

best on squared image patches of a power of two size.

3.2 Local Texton Dissimilarity

Algorithm

Algorithm 1 computes the LTD between gray-scale

texture images img

and img

, using the underlying

Bhattacharyya coefﬁcient to compute the similarity

between texton feature vectors.

Algorithm 1. Local Texton Dissimilarity

Input:

img

– a gray-scale texture image of h

× w

pixels;

img

– another gray-scale texture image of h

× w

pixels;

n – the number of features that represent a texton;

gridStep – the skip step that generates a dense grid

over the image;

offsetStep – the skip step used for comparing patches

at different offsets;

w – a vector of feature weights (some features can be

more important than others);

th – the texton similarity threshold.

Initialization:

dist = 0

h = min{h

} − p + 1

w = min{w

} − p + 1

Computation:

for x = 1:gridStep:h

for y = 1:gridStep:w

get texton

at position (x,y) in img

d = 0

while NO texton at offset d similar to texton

get texton

at offset d from (x, y) in img

∑

i=1



texton

− w

texton



if s

< th

dist = dist + d

break

endif

if all textons at offset d were tested

d = d+ offsetStep

endif

endwhile

get texton

at position (x,y) in img

d = 0

while NO texton at offset d similar to texton

get texton

at offset d from (x, y) in img

∑

i=1



texton

− w

texton



if s

< th

dist = dist + d

break

endif

if all textons at offset d were tested

d = d+ offsetStep

endif

endwhile

endfor

Output: dist – the dissimilarity between textures

img

and img

Algorithm 1 needs a few input parameters besides

the two images. The number of features gives the size

of the feature vector. In this work, the 9 features de-

scribed in Section 3.1 were used. In the algorithm,

texton

represents the i-th feature of the texton repre-

sentation, and w

represents the weight associated to

the i-th feature.

The results of the LTD algorithm can further be

improved by adding more features or probably by us-

ing completely different features. The parameter that

generates a dense grid over the image, and the skip

step used for comparing patches at different offsets

are used to speed up the LTD algorithm without loos-

ing too much accuracy. These parameters induce a

sparse representation of the images. Using a sparse

representation is indeed necessary, since patch-based

algorithms are heavy to compute with current com-

puters because they usually manipulate millions of

patches (Barnes et al., 2011). The texton similarity

threshold is a value in the [0, 1] interval, that deter-

mines when two textons are considered to be similar.

All these parameters need to be adjusted with regard

to the data set size and to the image dimensions, in or-

der to obtain a good trade-off between accuracy and

speed.

4 EXPERIMENTS

In the experiments, LTD is evaluated with different

kernel methods to show that good performance levels

are due to the use of LTD. Two data sets of texture im-

ages are used to assess the performance of several ker-

nel methods based on LTD, namely the Brodatz data

set and the UIUCTex data set. Another experiment is

performed to show the application of LTD on biomass

type identiﬁcation. All the experiments presented in

this work aim at showing that LTD has general ap-

plications for texture classiﬁcation, and that LTD is

indeed a robust dissimilarity measure.

4.1 Data Sets

The ﬁrst data set used for testing the dissimilarity

presented in this paper is the Brodatz data set (Bro-

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

596

datz, 1966). This data set is probably the best known

benchmark used for texture classiﬁcation, but also one

of the most difﬁcult, since it contains 111 classes with

only 9 samples per class. Samples of 213 × 213 pix-

els are cut using a 3 by 3 grid from larger images of

640 × 640 pixels. Figure 1 presents three sample im-

ages per class of three classes randomly selected from

the Brodatz data set.

Figure 1: Sample images from three classes of the Brodatz

data set.

The second experiment is conducted on the

UIUCTex data set of (Lazebnik et al., 2005b). It con-

tains 1000 texture images of 640 × 480 pixels rep-

resenting different types of textures such as bark,

wood, ﬂoor, water, and more. There are 25 classes

of 40 texture images per class. Textures are viewed

under signiﬁcant scale, viewpoint and illumination

changes. Images also include non-rigid deformations.

This data set is available for download at http://www-

cvr.ai.uiuc.edu/ponce grp.

The third experiment is conducted on a new data

set of biomass texture images provided by this work.

It contains 270 images of 512 × 512 pixels represent-

ing close up photos of three types of biomass resulted

after the processing of wheat, municipal waste and

corn, respectively. Photos where taken at different

zoom levels under various lighting conditions. Fig-

ure 2 show a few random samples of biomass im-

ages. There are 90 images per class. The goal is

to build a classiﬁer that is able to identify the three

types of biomass: wheat, waste, and corn, respec-

tively. The Biomass Texture data set is available for

use at http://biomass.herokuapp.com.

4.2 Learning Methods

To use LTD for texture classiﬁcation, it should be

plugged into a similarity-based learning method. Sev-

eral similarity-based classiﬁers are proposed. The

ﬁrst one is the Nearest Neighbors model (k-NN). It

was chosen because it directly reﬂects the discrimina-

tory power of the dissimilarity measure. Several state

of the art kernel methods are also used, namely the

Kernel Ridge Regression (KRR), the Support Vector

Machines (SVM), the Kernel Discriminant Analysis

Figure 2: Sample images from the Biomass Texture data

set.

(KDA), and the Kernel Partial Least Squares (KPLS).

Kernel methods are based on similarity. LTD can be

transformed into a similarity measure by using the

Gaussian-like kernel (also known as the RBF kernel):

k(img

,img

) = exp



−

LTD(img

,img

)

2σ



where img

and img

are two gray-scale texture im-

ages. The parameter σ is usually chosen to match the

number of features so that values of k(img

,img

) are

well scaled.

For a particular classiﬁcation problem, some ker-

nel methods may be more suitable than others. The

accuracy level depends on many aspects such as class

distribution, the number of classes, data noise, size

of the training data, and so on. For example, the

KRR classiﬁer can be used with success for problems

with well-balanced classes. But, in some particular

cases, when the number of classes is greater than 2,

there is a serious problem with the regression meth-

ods. More precisely, some classes can be masked by

others. The KDA classiﬁer is able to improve accu-

racy by avoiding the masking problem (Hastie and

Tibshirani, 2003).

4.3 Brodatz Experiment

The baseline method proposed for this experiment is

a 1-NN model that is based on the Bhattacharyya co-

efﬁcient computed on the 9 texture features described

in Section 3.1. The features are extracted from entire

images. The second proposed model is a 1-NN clas-

siﬁer based on LTD. The baseline is useful to assess

the performance gained by the use of LTD. The other

proposed classiﬁers are the KRR, the KPLS, the SVM

and the KDA, all based on LTD. The KDA method is

particularly suitable for problems with many classes,

such as Brodatz.

LocalTextonDissimilaritywithApplicationsonBiomassClassification

597

In (Lazebnik et al., 2005b), the accuracy rate re-

ported on the Brodatz data set using 3 training sam-

ples per class is 88.15%. Table 1 compares accuracy

rates of the proposed classiﬁers with the accuracy rate

of the state of the art method described in (Lazebnik

et al., 2005b), using the same setup with 3 random

samples per class for training. The accuracy rates pre-

sented in Table 1 are actually averages of accuracy

rates obtained over 20 runs for each method. The 1-

NN based on LTD model has a far better accuracy

than the baseline, proving that LTD helps the learning

method to achieve better results. All the kernel meth-

ods based on LTD are above the state of the art clas-

siﬁer. The best classiﬁer among them is KDA, which

has an accuracy of 90.87%. It is 5.46% better than the

1-NN based on LTD, and 2.72% better that the state of

the art method. It seems that LTD is a good dissimilar-

ity measure for texture classiﬁcation. Combined with

suitable learning methods, LTD gives results compa-

rable to state of the art method. Despite better tex-

ture classiﬁcation methods exist (Zhang et al., 2007),

the classiﬁers based on LTD can also be improved by

adding more features to the texton representation.

Table 1: Accuracy rates on the entire Brodatz data set using

3 random samples per class for training. Learning methods

based on LTD are compared with the state of the art method.

Method Accuracy

baseline 1-NN 77.68%

Best of (Lazebnik et al., 2005b) 88.15%

1-NN + LTD 85.41%

KRR + LTD 89.43%

SVM + LTD 89.48%

KPLS + LTD 89.57%

KDA + LTD 90.87%

In this experiment, LTD was computed on patches

of 32 × 32 pixels, using a similarity threshold of 0.02

and a maximum offset of 80 pixels. Patches were ex-

tracted on a dense grid with a gap of 32 pixels. Fea-

ture weighting can improve accuracy by almost 1%.

Thus, adjusting feature weights is not very impor-

tant, but it helps the classiﬁer. However, the feature

weights were manually adjusted to increase the im-

portance of Gabor features and fractal dimension by

a factor of two, and to decrease the importance of the

mean and the standard deviation by a factor of two.

The weights were tuned on the baseline 1-NN model,

which also uses feature weighting in the reported re-

sults. The parameter σ of the LTD kernel was cho-

sen to be 10

−3

. All the parameters were chosen by

cross validation on a subset of the Brodatz data set.

An interesting remark is that these parameters do not

change by too much on the other data sets.

Using these parameters, it takes less than 1 second

to compute LTD between two images on a computer

with Intel Core Duo 2.26 GHz processor and 4 GB

of RAM memory using a single Core. Reported ac-

curacy rates can be improved by a few percents using

a more dense grid and a greater maximum offset, but

the LTD computation will also take more time. How-

ever, with the current parameters, LTD is much faster

than Local Patch Dissimilarity, which takes about 5

minutes to compare two images from the Brodatz data

set with similar parameters, without skipping overlap-

ping patches.

4.4 UIUCTex Experiment

In this experiment, the same classiﬁers evaluated on

the Brodatz data set are also evaluated on the UIUC-

Tex data set. More precisely, the evaluated classi-

ﬁers are the baseline 1-NN model based on the Bhat-

tacharyya coefﬁcient, the 1-NN classiﬁer based on

LTD, and the kernel classiﬁers based on LTD, namely

the KRR, the KPLS, the SVM, and the KDA. These

classiﬁers are compared with the state of the art clas-

siﬁer of (Lazebnik et al., 2005b). The best accuracy

level of the state of the art classiﬁer on the UIUCTex

data set, reported in (Lazebnik et al., 2005b) using 20

training samples per class, is 97.41%.

Table 2 compares accuracy rates of the classiﬁers

based on LTD with the accuracy rate of the state of

the art classiﬁer of (Lazebnik et al., 2005b), using

the same setup with 20 random samples per class for

training. The accuracy rates are averaged over 20

runs for each method. The accuracy of the 1-NN

model based on LTD is 9.32% better than accuracy of

the baseline 1-NN, proving again that LTD is able to

achieve much better results. However, the accuracy of

the 1-NN based on LTD is far behind the state of the

art classiﬁer. Even the kernel methods have accuracy

rates that are roughly 4% lower than the state of the

art classiﬁer. The best classiﬁer based on LTD is the

KPLS, with an accuracy of 93.79%, which is 3.62%

lower than the state of the art method. The accuracy

of these kernel methods depend on LTD, which de-

pends in turn on the features extracted from images to

obtain textons. Better features will result in a dissim-

ilarity measure capable of making ﬁner distinctions,

and, consequently, in a better kernel classiﬁer. But

even with the 9 features proposed in Section 3.1, LTD

seems to give results that are comparable to the state

of the art method.

In this experiment, LTD was computed on patches

of 64 × 64 pixels, using a similarity threshold of 0.02

and a maximum offset of 240 pixels. Patches were

extracted on a dense grid with a gap of 64 pixels. The

same feature weights as in the Brodatz experiment

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

598

Table 2: Accuracy rates on the UIUCTex data set using 20

random samples per class for training. Learning methods

based on LTD are compared with state of the art method.

Method Accuracy

baseline 1-NN 79.34%

Best of (Lazebnik et al., 2005b) 97.41%

1-NN + LTD 88.66%

KRR + LTD 93.51%

SVM + LTD 93.62%

KPLS + LTD 93.79%

KDA + LTD 93.38%

were used. The parameter σ of the LTD kernel was

chosen to be 10

−3

. All the parameters were chosen

by cross validation on a subset of the UIUCTex data

set.

4.5 Biomass Experiment

The classiﬁers evaluated in this experiment are the

baseline 1-NN model based on the Bhattacharyya co-

efﬁcient, the 1-NN classiﬁer based on LTD, and the

kernel classiﬁers based on LTD, namely the KRR,

the KPLS, the SVM, and the KDA. These classiﬁers

must identify the three classes of biomass from the

Biomass Texture data set.

Table 3 presents accuracy rates of the proposed

classiﬁers using three different setup procedures. The

ﬁrst setup is to use 20 random samples per class for

training and the rest of 70 samples for testing. The

second setup is to use 30 random samples per class

for training and 60 samples for testing. The last setup

is to use 40 random samples per class for training and

50 samples for testing. The accuracy rates are aver-

aged over 50 runs for each method. As expected, the

accuracy of each method improves when more train-

ing samples are used. For example, the accuracy of

the baseline method grows by 6.83% from 20 training

samples to 40 training samples. However, the classi-

ﬁers based on LTD are more stable, since the accu-

racy of each classiﬁer grows only by roughly 3 − 4%

from 20 training samples to 40 samples. The learning

methods based on LTD show a signiﬁcant improve-

ment in accuracy over the baseline. The best classi-

ﬁer based on LTD is KPLS. In all the test cases, the

KPLS based on LTD has an accuracy of at least 10%

better than the accuracy of the baseline 1-NN. Over-

all, the kernel classiﬁers achieve roughly similar ac-

curacy levels. The empirical results show again that

LTD is a powerful dissimilarity measure for texture

classiﬁcation.

In this experiment, LTD was computed on patches

of 64 × 64 pixels, using a similarity threshold of 0.02

and a maximum offset of 256 pixels. Patches were ex-

tracted on a dense grid with a gap of 64 pixels. Again,

Table 3: Accuracy rates on Biomass Texture data set using

20, 30 and 40 random samples per class for training and 70,

60 and 50 for testing, respectively.

Method 20/70 30/60 40/50 Acc.

baseline 1-NN 80.35% 84.72% 87.18%

1-NN + LTD 88.09% 90.20% 91.28%

KRR + LTD 93.72% 96.40% 97.64%

SVM + LTD 93.98% 96.58% 97.72%

KPLS + LTD 94.48% 96.90% 97.97%

KDA + LTD 94.08% 96.40% 97.67%

feature weights were adjusted to increase the impor-

tance of Gabor features and fractal dimension by a

factor of two, and to decrease the importance of the

mean and the standard deviation by a factor of two.

The parameter σ of the LTD kernel was chosen to be

−3

. All the parameters were chosen by cross vali-

dation on a subset of the Biomass Texture data set.

5 CONCLUSIONS AND FURTHER

WORK

This work presented a texture dissimilarity measure

based on textons, called Local Texton Dissimilarity.

It is based on the idea of comparing textons that are

represented as a set of features extracted from image

patches. Experiments showed that LTD can be used

to obtain accuracy levels comparable to state of the art

methods. The proposed dissimilarity measure showed

its application on biomass type identiﬁcation. To as-

sess the performance level of LTD on biomass clas-

siﬁcation, a new data set of biomass texture images

was provided by this work. On this data set, the accu-

racy level of a classiﬁer based on LTD can be as high

as 97.97%, which is more than enough for a practical

application.

In future work, LTD can be improved by adding

more features to the texton feature set, or by chang-

ing the features completely. For example, textons can

be obtained by vector quantizing local image descrip-

tors, such as the SIFT descriptor (Lowe, 1999). Fi-

nally, a system for biomass type identiﬁcation will be

designed to analyze photos taken on mobile devices.

A classiﬁer based on LTD will be integrated in this

system.

ACKNOWLEDGEMENTS

The contribution of the authors to this paper is equal.

LocalTextonDissimilaritywithApplicationsonBiomassClassification

599

REFERENCES

Barnes, C., Goldman, D. B., Shechtman, E., and Finkel-

stein, A. (2011). The PatchMatch Randomized Match-

ing Algorithm for Image Manipulation. Communica-

tions of the ACM, 54(11):103–110.

Brodatz, P. (1966). Textures: a photographic album for

artists and designers. Dover pictorial archives. Dover

Publications, New York, USA.

Dash, J., Mathur, A., Foody, G. M., Curran, P. J., Chip-

man, J. W., and Lillesand, T. M. (2007). Land cover

classiﬁcation using multi-temporal MERIS vegetation

indices. IJRS, 28(6):1137–1159.

Daugman, J. G. (1985). Uncertainty relation for resolution

in space, spatial frequency, and orientation optimized

by two-dimensional visual cortical ﬁlters. Journal of

the Optical Society of America A, 2(7):1160–1169.

Dinu, L. P., Ionescu, R., and Popescu, M. (2012). Lo-

cal Patch Dissimilarity for Images. Proceedings of

ICONIP, 7663:117–126.

Efros, A. A. and Freeman, W. T. (2001). Image quilting for

texture synthesis and transfer. Proceedings of SIG-

GRAPH ’01, pages 341–346.

Falconer, K. (2003). Fractal Geometry: Mathematical

Foundations and Applications. Wiley, 2 edition.

Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973).

Textural Features for Image Classiﬁcation. IEEE

Transactions on Systems, Man and Cybernetics,

3(6):610–621.

Hastie, T. and Tibshirani, R. (2003). The Elements of Sta-

tistical Learning. Springer, corrected edition.

Hoekman, D. H. and Quinnones, M. J. (2000). Land cover

type and biomass classiﬁcation using AirSAR data fro

evaluation of monitoring scenarios in the Colombian

Amazon. IEEE Transactions on Geoscience and Re-

mote Sensing, 38:685–696.

Kuse, M., Wang, Y.-F., Kalasannavar, V., Khan, M., and

Rajpoot, N. (2011). Local isotropic phase symmetry

measure for detection of beta cells and lymphocytes.

Journal of Pathology Informatics, 2(2):2.

Lazebnik, S., Schmid, C., and Ponce, J. (2005a). A Maxi-

mum Entropy Framework for Part-Based Texture and

Object Recognition. Proceedings of ICCV, 1:832–

838.

Lazebnik, S., Schmid, C., and Ponce, J. (2005b). A Sparse

Texture Representation Using Local Afﬁne Regions.

PAMI, 27(8):1265–1278.

Leung, T. and Malik, J. (2001). Representing and Recogniz-

ing the Visual Appearance of Materials using Three-

dimensional Textons. IJCV, 43(1):29–44.

Lowe, D. G. (1999). Object Recognition from Local Scale-

Invariant Features. Proceedings of ICCV, 2:1150–

1157.

Popescu, A. L., Popescu, D., Ionescu, R. T., Angelescu, N.,

and Cojocaru, R. (2013). Efﬁcient Fractal Method for

Texture Classiﬁcation. Proceedings of ICSCS, (Ac-

cepted).

Varma, M. and Zisserman, A. (2005). A Statistical Ap-

proach to Texture Classiﬁcation from Single Images.

IJCV, 62(1-2):61–81.

Wulder, M. A., White, J. C., Fournier, R. A., Luther, J. E.,

and Magnussen, S. (2008). Spatially Explicit Large

Area Biomass Estimation: Three Approaches Using

Forest Inventory and Remotely Sensed Imagery in a

GIS. Sensors, 8(1):529–560.

Xie, J., Zhang, L., You, J., and Zhang, D. (2010). Texture

classiﬁcation via patch-based sparse texton learning.

Proceedings of ICIP, pages 2737–2740.

Zhang, J., Marszalek, M., Lazebnik, S., and Schmid, C.

(2007). Local Features and Kernels for Classiﬁcation

of Texture and Object Categories: A Comprehensive

Study. IJCV, 73(2):213–238.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

600