CONTENT-BASED TEXTURE IMAGE RETRIEVAL USING THE

LEMPEL-ZIV-WELCH ALGORITHM

Leonardo Vidal Batista, Moab Mariz Meira

Departamento de Inform

atica - DI / UFPB

Cidade Universit

aria - 58.051-900

ao Pessoa - PB - Brasil

Nicomedes L. Cavalcanti J

unior

Centro de Inform

atica - CIn / UFPE

Av. Prof. Luiz Freire, s/n, Cidade Universit

aria - 50.740-540

Recife - PE - Brasil

Keywords:

Content-based Image Retrieval, Texture classiﬁcation, Lempel-Ziv-Welch.

Abstract:

This paper presents a method for content-based texture image retrieval using the Lempel-Ziv-Welch (LZW)

compression algorithm. Each texture image in the database is processed by a global histogram equalization

ﬁlter, and then an LZW dictionary is constructed for the ﬁltered texture and stored in the database. The

LZW dictionaries thus constructed comprise a statistical model to the texture. In the query stage, each texture

sample to be searched is processed by the histogram equalization ﬁlter and successively encoded by the LZW

algorithm in static mode, using the stored dictionaries. The system retrieves a ranked list of images, sorted

according to the coding rate achieved with each stored dictionary. Empirical results with textures from the

Brodatz album show that the method achieves retrieval accuracy close to 100%.

1 INTRODUCTION

Keyword annotation is the most traditional image re-

trieval paradigm. In this approach, the images are ﬁrst

annotated manually by keywords. They can then be

retrieved by their corresponding annotations. How-

ever, there are three main difﬁculties with this ap-

proach, i.e., the large amount of manual effort re-

quired in developing the annotations, the differences

in interpretation of image contents, and the incon-

sistency of the keyword assignments among differ-

ent indexers (Faloutsos et al., 1993; Flickner et al.,

1995). As the size of image repositories increases,

the keyword annotation approach becomes infeasible.

To overcome the difﬁculties of the keyword-based ap-

proach, an alternative mechanism, content-based im-

age retrieval (CBIR) was proposed in the early 1990’s.

CBIR consists in using visual features, which are im-

age primitives, such as color, texture, and shape fea-

tures, as the image index. This approach has the ad-

vantage of automatic feature extraction.

The explosive growth of digital image technolo-

gies in the last years brings the necessity to investi-

gate and develop new search tools to efﬁciently lo-

cate pictorial information. Conventional search tools

generally allow exclusively textual queries. Most In-

ternet search tools speciﬁcally designed for images

looks for textual relevant information about image

content by analyzing the ﬁlename of the graphic ﬁles,

’META’ tags and ALT attributes of the ’IMG’ tags

in the HTML code of the pages, terms near the im-

ages in the pages, nature and orientation of sites and

providers, etc. Conventional image databases gener-

ally stores textual information along with the images,

thus allowing textual queries. However, it is known

that textual queries, notably in the scope of multime-

dia databases, tend to present a great number of irrele-

vant results while failing to present many relevant re-

sults. Many efforts have been made to develop CBIR

tools, by extracting and analyzing pictorial features

such as shapes, colors and textures (Zibreira, 2000).

CBIR has attracted great research attention, rang-

ing from government (Jain, 1993; Jain et al., 1995)

and industry (Bach et al., 1996; Dowe, 1993; Flickner

et al., 1995), to universities (Huang et al., 1996; Ma

and Manjunath, 1999; Mandal et al., 1997; Pentland

et al., 1996; Smith and Chang, 1997). Even ISO/IEC

has deﬁned MPEG-7 (ISO/IEC et al., 1997a; ISO/IEC

et al., 1997b; ISO/IEC et al., 1997c) to encompass

a standard multimedia content description interface.

Many CBIR systems, both commercial (Bach et al.,

1996; Dowe, 1993; Faloutsos et al., 1993; Flickner

et al., 1995) and academic have been developed re-

cently.

Vidal Batista L., Mariz Meira M. and L. Cavalcanti Júnior N. (2006).

CONTENT-BASED TEXTURE IMAGE RETRIEVAL USING THE LEMPEL-ZIV-WELCH ALGORITHM.

In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 62-68

DOI: 10.5220/0001376600620068

 SciTePress

In Flickner’s and Addi’s works (Flickner et al.,

1995; Addis et al., 2002), tools which approach some

of the relevant aspects of CBIR are presented. These

tools were applied to construct a museum and gallery

web applications respectively which allow search for

artworks by selecting predominant colours from a

palette or by sketching shapes on a canvas.

Texture is a fundamental attribute used by the

human visual system and computer vision systems

for segmentation, classiﬁcation and interpretation of

scenes (Porat and Zeevi, 1989). There has been a

great interest in the development of texture-based pat-

tern recognition methods in different areas, such as

remote sensing (Dell’Acqua and Gamba, 2003; Au-

gusteijn et al., 1995), image-based medical diagnosis

(Southard and Southard, 1996), industrial automation

(Kumar and Pang, 2002) and biometric recognition

(Jain et al., 2004; He et al., 2004).

Although intuitively recognized by the human vi-

sual system, texture is not easy to characterize for-

mally. The problem resides in the intrinsic difﬁculty

to deﬁne what is most relevant to characterize texture,

as the answer depends on subjective perceptual con-

siderations and on particular applications. Therefore,

texture feature extraction and modeling tends to be a

difﬁcult and application-driven task. A popular yet

rather vague deﬁnition states that textures are spatial

patterns formed by more or less accurate repetitions

of some basic subpatterns (Baheerathan et al., 1999).

Modern lossless data compression algorithms have

been applied to pattern recognition problems, due

to their ability to construct accurate statistical mod-

els, in some cases with low computational require-

ments (Bell et al., 1990). A solid theoretical foun-

dation for using LZ78 (Ziv and Lempel, 1978) and

other dictionary-based compression algorithms (Bell

et al., 1990) for pattern classiﬁcation is well estab-

lished (Ziv, 1988; Ziv and Merhav, 1993).

A potential problem associated with lossless

dictionary-based compression algorithms for image

retrieval is the fact that these methods search exact

matches in the dictionary for strings in the message

to be compressed. A precise dictionary constructed

for a given texture class may present a poor per-

formance when compressing a new sample from the

same class, if this new sample presents gray-level de-

viations caused by digitization noise or illumination

changes.

Degraded performance, even when gray-level devi-

ations are subtle, indicates that the constructed model

may not be able to adequately describe the new tex-

ture sample, and consequently classiﬁcation accuracy

may also degrade.

Two possible solutions to this problem are:

1. Adoption of a lossy dictionary-based compressor

(Finamore and Leister, 1996), less sensitive to

small, spurious gray-level deviations;

2. Reduction of these deviations by means of image

processing techniques, prior to the use of a lossless

dictionary-based compression algorithm.

Histogram equalization is a well-known non-linear

operation that generates an approximately uniform

distribution of gray-levels over the available range

(Bovik, 2000). Histogram equalization tends to map

to the same value multiple gray levels that have sim-

ilar values, thus reducing the small gray level devia-

tions that tends to cause mismatches in the searching

stage of lossless dictionary-based compressors. The

operation also decreases the probability that an im-

age retrieval system discriminates texture classes by

average gray level or variance, instead of by its tex-

tural properties (Randen and Husy, 1999). This al-

lows a more precise evaluation of the capabilities of

the method to discriminate texture attributes.

Image retrieval based on universal data compres-

sion models have several potential advantages over

classical methods: since there is no feature selection,

no information is discarded - the models describe the

classes as a whole (Frank et al., 2000); no assump-

tions about the probability distributions of the classes

are required; the adaptive model construction capabil-

ity of compression algorithms offers a uniform way

to work with different types of sources (Ojala et al.,

2002); the similarity rule is very simple (Teahan and

Harper, 2001).

This paper proposes a new content-based image

retrieval method for texture images using histogram

equalization and the Lempel-Ziv-Welch (LZW) loss-

less compression algorithm (Welch, 1984). The

rest of this paper is organized as follows. Sec-

tion 2 presents some fundamental concepts; section 3

presents the LZW algorithm; section 4 describes the

proposed method; section 5 presents the empirical

evaluation of the proposed search tool; and section 6

presents a discussion of the results and the concluding

remarks.

2 ENTROPY AND MARKOV

MODELS

Let S be a stationary discrete information source

that generates messages over a ﬁnite alphabet A =

,...,a

}. The source chooses successive

symbols from A according to some probability dis-

tribution that, in general, depends on preceding sym-

bols. A generic message is modeled as a station-

ary stochastic process x =

...

−2

−1

2...

, with

∈ A. Let x

= x

...x

represent a mes-

sage of length n. Since |A|= M, the source can gen-

erate M

different messages of length n. Let x

i =1, 2,...,M

denote the i

of these messages,

according to some sorting order, and assume that the

CONTENT-BASED TEXTURE IMAGE RETRIEVAL USING THE LEMPEL-ZIV-WELCH ALGORITHM

source follows a probability distribution P , so that

message x

is produced with probability P (x

Let

(P )=−



i=1

P (x

) log

P (x

) (1)

(P ) decreases monotonically with n (Shannon,

1948) and the entropy of the source is:

H(P ) = lim

n→∞

(P ) bits/symbol. (2)

An alternative formulation for H(P ) uses condi-

tional probabilities. Let P (x

n−1

) be the prob-

ability of sequence x

= x

n−1

concate-

nated with x

= a

) and let P (a

n−1

P (x

n−1

)P (x

n−1

) be the probability of x

= a

conditioned on x

n−1

. The entropy of the n

order

approximation to H(P ) (Shannon, 1948) is:

(P )=−



i=1



j=1

P (x

n−1

) log

P (a

n−1

)

bits/symbol.

(3)

(P ) decreases monotonically with n (Shannon,

1948) and the entropy of the source is:

H(P ) = lim

n→∞

(P ) bits/symbol. (4)

Eq. 4 involves the estimation of probabilities con-

ditioned on an inﬁnite sequence of previous symbols.

When ﬁnite memory is assumed the sources can be

modeled by a Markov process of order n − 1, so that

P (a

| ...x

n−2

n−1

)=P (a

...x

n−1

). In this

case, H(P )=F

(P ).

Deﬁne the coding rate of a coding scheme as the

average number of bits per symbol the scheme uses

to encode the source output. A lossless compressor

is a uniquely decodable coding scheme whose goal

is to achieve a coding rate as small as possible. The

coding rate of any uniquely decodable coding scheme

is always greater than or equal to the source entropy

(Shannon, 1948). Optimum coding schemes have

a coding rate equal to the theoretical lower bound

H(P ), thus achieving maximum compression.

For Markov processes of order n − 1, optimum en-

coding is reached if and only if symbol x

= a

oc-

curring after x

n−1

is coded with − log

P (a

n−1

)

bits (Bell et al., 1990; Shannon, 1948). However, it

may be impossible to accurately estimate the condi-

tional distribution P (.|x

n−1

) for large values of n,

due to the exponential growth of the number of dif-

ferent contexts, which brings well-known problems,

such as context dilution (Bell et al., 1990).

3 THE LZW ALGORITHM

Even though the source model P is generally un-

known, it is possible to construct a coding scheme

based upon some (possibly implicit) probabilistic

model Q that approximates P . The better Q approx-

imates P , the smaller the coding rate achieved by the

coding scheme.

In order to achieve low coding rates, modern loss-

less compressors rely on the construction of sophis-

ticated models that closely follows the true source

model. Statistical compressors encode messages

according to an estimated statistical model for the

source. Dictionary-based compressors replace strings

of symbols from the message to be encoded with in-

dexes into a dictionary of strings, which is generally

adaptively constructed during the encoding process.

When greedy parsing is used, at each step the en-

coder searches the current dictionary for the longest

string that matches the next sequence of symbols in

the message, and replaces this sequence with the in-

dex of the longest matching string in the dictionary.

Dictionary-based compressors with greedy parsing,

such as LZW, are highly popular because they com-

bine computational efﬁciency with low coding rates.

It has been proved that each dictionary-based com-

pressor with greedy parsing has an equivalent statisti-

cal coder that achieves the same compression (Bell

et al., 1990). In dictionary-based coding, the dic-

tionary embeds an implicit statistical model for the

source.

The initial LZW dictionary contains all possible

strings of length one. The LZW algorithm ﬁnds the

longest string, starting from the ﬁrst symbol of the

message, which is already present in the dictionary.

This string is coded with the index for the matching

string in the dictionary, and the string is extended with

the next symbol in the message, x

. The extended

string is added to the dictionary and the process re-

peats, starting from x

(Bell et al., 1990).

LZW achieves optimum asymptotic performance

for Markov sources, in the sense that its coding rate

tends to the entropy of the source as the length of the

message to be coded tends to inﬁnity (Savari, 1997).

It means that LZW algorithm learns a progressively

better model for the source during encoding, and a

perfect model for the source is learned when an in-

ﬁnite message has been coded. In practice, since

the messages to be compressed are ﬁnite, LZW only

learns an approximate model for the source.

4 THE PROPOSED METHOD

The accurate models built by modern lossless com-

pressors can be used to characterize texture. Any efﬁ-

VISAPP 2006 - IMAGE UNDERSTANDING

cient model-based lossless compressor could be used,

but LZW algorithm was chosen due to its good com-

promise between coding efﬁciency and computational

requirements (Bell et al., 1990).

4.1 Model Learning and Storage

Consider a database containing N texture samples

, i =1, 2,...,N. The samples are n x n im-

ages extracted from histogram-equalized images. In

the model learning stage, the LZW algorithm com-

presses sample t

following the horizontal scanning

order shown in Figure 1.a, and the resulting dictio-

nary H

is stored in the database as a model for the

horizontal structure of t

, i =1, 2,...,N. The LZW

algorithm then compresses t

following the vertical

scanning order shown in Figure 1.b, and the resulting

dictionary V

is stored in the database as a model for

the vertical structure of T

, i =1, 2,...,N.

Figure 1: Scanning orders. (a) Horizontal; and (b) Vertical.

4.2 The Retrieval Stage

In the retrieval stage, LZW operates in static mode.

In this mode, one of the dictionaries generated in the

model learning stage is used to encode a query sam-

ple, and no new strings are added to the dictionary

during the encoding process.

A n x n query sample x is coded by the LZW algo-

rithm with static dictionary H

, following the horizon-

tal scanning order shown in Figure 1.a., and the corre-

sponding coding rate h

is registered, i =1, 2,...,N.

Then the LZW algorithm with static dictionary V

en-

codes x, following the vertical scanning order shown

in Figure 1.b, and the corresponding coding rate v

registered, i =1, 2,...,N. As in the previous stage,

all samples are extracted from histogram-equalized

images.

Let

+ v

(5)

Query sample x is considered more similar to tex-

ture t

than to texture t

in the database if r

< r

i, j =1, 2,...,N. The rationale is that if x is more

similar to t

than to any other texture in the database

(according to the texture models), the dictionaries H

and V

embeds the model that best describes its hori-

zontal and vertical structure, thus yielding the small-

est coding rates.

5 EXPERIMENTAL RESULTS

The complete Brodatz album (Brodatz, 1966), ob-

tained from a public archive, was used to evaluate the

performance of the proposed method. All 112 tex-

tures have 640 x 640 pixels, with 8 bits/pixel. In the

experiments, each Brodatz texture is taken as a sin-

gle class. This corpus is the same used by (Xu et al.,

2000), thus allowing direct comparison with another

CBIR system from the literature. Some examples of

the Brodatz textures are shown in Figure 2. Notice

that some of Brodatz images are so irregular that they

cannot be considered as texture images, according to

most accepted deﬁnitions for texture. These highly ir-

regular images have a negative effect on recognition

rate (Xu et al., 2000).

In order to assess the effect of histogram equaliza-

tion, tests were made with and without applying this

operation before model learning and retrieval. In this

section and in the next one, the proposed method with

and without histogram equalization will be identi-

ﬁed as CBIR-LZW-HE and CBIR-LZW, respectively.

Each Brodatz texture was partitioned in 128 x 128

non-overlapping subimages, which were taken as tex-

ture samples. Nine of these samples were stored in the

database along with the corresponding LZW dictio-

naries constructed as describe in section 4.1, and the

other ones were used as query samples for testing the

accuracy of the method. Therefore, the database con-

tains 112 texture classes, hereafter identiﬁed as C

i =0, 1,...,111, and each class C

has nine samples,

identiﬁed as s

, i =0, 1,...,111,j =0, 1,...,8,

for a total of 9 x 112 = 1008 texture samples. Given

a query sample x from class C

, retrieval of any one of

the nine samples s

,j =0, 1,...,8, from the same

class is considered successful.

A number of accuracy measures are used in the lit-

erature. In order to directly compare the results of the

proposed method with those of (Xu et al., 2000), the

adopted accuracy measure was the average recogni-

tion rate. Consider that for each query the system re-

turns the c most similar texture samples from the data-

base, according to the LZW models. For a query sam-

ple from class C

, let h(c) be the number of samples

from C

in the c retrieved samples. Recognition rate,

R(c), is then deﬁned as the ratio between h(c) and the

total number of samples from classes C

stored in the

database:

R(c)=

h(c)

(6)

The average recognition rate, AR(c), is the average

value of R(c) for all test queries. From the deﬁnition,

it follows that AR(c) does not decrease as c increases

and, since there are 1008 texture samples in the test

database, AR(c)=1for c = 1008. More efﬁcient re-

trieval mechanisms tends to achieve AR(c)=1for

CONTENT-BASED TEXTURE IMAGE RETRIEVAL USING THE LEMPEL-ZIV-WELCH ALGORITHM

smaller values of c than less efﬁcient mechanisms.

In the ideal case for a database with 9 samples/class,

AR(9) = 1.

Results are shown in Table 1. Figure 2 presents as

an example a query which resulted in a low AR(9),

and Figure 3 presents the whole Brodatz texture from

which this query samples was taken. Figure 4 com-

pares the performance of the proposed method with

M2, which achieved the best performance among the

four variations of the method proposed by (Xu et al.,

2000).

Figure 2: A query resulting in AR(9) = 0.3333 with

CBIR-LZW-HE. Sample 0 is the query sample, sample 1 is

the ﬁrst retrieved sample, sample 2 is the second retrieved

sample, and so on.

Figure 3: Texture D38, from which the query sample in

Figure 2 was taken.

Table 1: AR(c) achieved by the proposed method, with and

without histogram equalization, for various c (number of

retrievals).

Sample size (n x n) CCR(%)

CLZW CLZW-GHE

4 x 4 80.6 99.9

8 x 8 97.9 100

16 x 16 99.7 100

32 x 32 100.0 100

Figure 4: AR(c) achieved by the proposed method, with

and without histogram equalization, and by the M2 method

reported in literature.

6 DISCUSSION AND

CONCLUSIONS

This paper proposed a new, simple and accurate

content-based image retrieval method for texture im-

ages using histogram equalization and the LZW algo-

rithm.

Table 1 shows that histogram equalization has a

strong positive impact on performance. CBIR-LZW-

HE achieved AR(9) = 0.9939(99.39%), while CBIR-

LZW achieved AR(40) = 0.9459(94.59%).

Figure 4 compares the performance of the proposed

method with M2 (Xu et al., 2000), over the same data-

base. The comparison shows the superiority of CBIR-

LZW-HE over M2. This superiority is still more re-

markable in the most restrictive case (c =9), with

CBIR-LZW-HE achieving AR(9) approximately 15

percent points above the value achieved by M2.

As stated before, many Brodatz images are so irreg-

ular that they cannot be considered as texture images,

according to accepted deﬁnitions for texture. Xu’s

work(Xu et al., 2000) assessed the effects of this pe-

culiarity of Brodatz album by dividing the textures in

VISAPP 2006 - IMAGE UNDERSTANDING

two separate databases, one containing only visually

homogeneous textures and the other one containing

only visually inhomogeneous textures. By analyzing

recognition rate over these two databases, they con-

cluded that irregular images have a negative effect on

recognition rate. On the other hand, CBIR-LZW-HE

results, with AR(c) near 100%, even for small values

of c, indicate that the method was to some extent ro-

bust to these irregularities.

Figure 2 presents an example query with a particu-

larly low AR(9)(33, 33%). As can be seen in Figure 3,

Brodatz texture D38, from which the query sample

was taken, presents a strong non-uniform illumination

variation, to the point of saturation in the lower right

corner, making recognition a difﬁcult task.

Future directions of research include assessing the

robustness of the method to gray-scale, rotation and

spatial-scale changes and investigating the use of

lossy dictionary-based compressors. In fact, prelim-

inary tests in this direction are already being con-

ducted. First results indicate that histogram equaliza-

tion makes the method robust to uniform gray-scale

variations. Some developments are also being inves-

tigated in order to make the method invariant to ro-

tation and non-uniform illumination variations, with

promising results. It should be noticed that although

invariance is an important feature in CBIR tools, in

many practical applications (e.g. in industrial quality

control by computer vision) images are acquired un-

der strictly controlled conditions, and practically do

not present scale, rotation or illumination changes.

The robustness of CBIR-LZW-HE to irregulari-

ties in the test database suggests that investigating

the applicability of the method to content based non-

textured images is also a very promising direction of

research.

REFERENCES

Addis, M., Lewis, P., and Martinez, K. (2002). Artiste im-

age retrieval system puts european galleries in the pic-

ture. Cultivate Interactive, (7).

Augusteijn, M. F., Clemens, L. E., and Shaw, K. A.

(1995). Performance evaluation of texture measures

for ground cover identiﬁcation in satellite images by

means of a neural network classiﬁer. IEEE Transac-

tions on Geoscience and Remote Sensing, 33(3):616–

626.

Bach, J. R., Fuller, C., Gupta, A., Hampapur, A.,

B. Horowitz, R. H., Jain, R., and Shu, C. F. (1996).

The virage image search engine: An open framework

for image management. In Proc. SPIE Storage and

Retrieval for Image and Video Databases IV, San Jose,

CA, USA.

Baheerathan, S., Albregtsen, F., and Danielsen, H. E.

(1999). New texture features based on the complexity

curve. Pattern Recognition, 32(4):605–618.

Bell, T. C., Cleary, J. G., and Witten, J. H. (1990). Text

Compression. Englewood Cliffs: Prentice-Hall.

Bovik, A. e. (2000). Handbook of Image and Video Process-

ing. San Diego: Academic Press.

Brodatz, P. (1966). Textures: A Photographic Album for

Artists and Designers. New York: Dover.

Dell’Acqua, F. and Gamba, P. (2003). Texture-based char-

acterization of urban environments on satellite sar im-

ages. IEEE Transactions on Geoscience and Remote

Sensing, 41(1):153–159.

Dowe, J. (1993). Content-based retrieval in multimedia

imaging. In Proc. SPIE Storage and Retrieval for Im-

age and Video Databases.

Faloutsos, C., Flickner, M., Niblack, W., Petkovic, D., Eq-

uitz, W., and Barber, R. (1993). Efﬁcient and effective

querying by image content. Tech. Rep., IBM Res. Rep.

Finamore, W. and Leister, M. A. (1996). Lossy lempel-ziv

algorithm for large alphabet sources and applications

to image compression. volume 1, pages 235–238.

Flickner, M., Sawhney, H., Niblack, W., Ashley, J., Huang,

Q., Dom, B., Gorkani, M., Haﬁne, J., Lee, D.,

Petkovic, D., Steele, D., and Yanker, P. (1995). Query

by image and video content: The qbic system. IEEE

Computer.

Frank, E., Chui, C., and Witten, I. H. (2000). Text cat-

egorization using compression models. In Storer,

J. A. and Cohn, M., editors, Proceedings of DCC-

00, IEEE Data Compression Conference, pages 200–

209, Snowbird, US. IEEE Computer Society Press,

Los Alamitos, US.

He, X., Hu, Y., Zhang, H., Li, M., Cheng, Q., and Yan, S.

(2004). Bayesian shape localization for face recogni-

tion using global and local textures. IEEE Transac-

tions on Circuits and Systems for Video Technology,

14(1):102–113.

Huang, T., Mehrotra, S., and Ramchandran, K. (1996).

Multimedia analysis and retrieval system (mars)

project. In Proc. 33rd Annu. Clinic Library Appl. of

Data Processing Digital Image Access and Retrieval.

ISO/IEC, JTC1/SC29/WG11, and N1920 (1997a). Mpeg-7:

Context and objectives (v.5).

ISO/IEC, JTC1/SC29/WG11, and N1921 (1997b). Third

draft of mpeg-7 requirements.

ISO/IEC, JTC1/SC29/WG11, and N1922 (1997c). Mpeg-7

applications document.

Jain, A. K., Ross, A., and Prabhakar, S. (2004). An in-

troduction to biometric recognition. IEEE Transac-

tions on Circuits and Systems for Video Technology,

14(1):4–20.

Jain, R. (1993). Nsf workshop on visual information man-

agement systems. SIGMOD Rec., 22(3):57–75.

Jain, R., Pentland, A., and Petkovic, D. (1995). NSF-ARPA

workshop visual inform. management syst.

CONTENT-BASED TEXTURE IMAGE RETRIEVAL USING THE LEMPEL-ZIV-WELCH ALGORITHM

Kumar, A. and Pang, G. K. H. (2002). Defect detection

in textured materials using optimized ﬁlters systems.

IEEE Transactions on Man and Cybernetics, Part B:

Cybernetics, 32(5):553–570.

Ma, W.-Y. and Manjunath, B. S. (1999). Netra: A tool-

box for navigating large image databases. Multimedia

Systems, 7(3):184–198.

Mandal, M. K., Panchanathan, S., and Aboulnasr, T. (1997).

Image indexing using translation and scale-invariant

moments and wavelets. In Storage and Retrieval for

Image and Video Databases (SPIE), pages 380–389.

Ojala, T., Pietikainen, M., and Maenpaa, T. (2002). Mul-

tiresolution gray-scale and rotation invariant texture

classiﬁcation with local binary patterns. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

24(7):971–987.

Pentland, A., Picard, R. W., and Sclaroff, S. (1996). Pho-

tobook: content-based manipulation of image data-

bases. Int. J. Comput. Vision, 18(3):233–254.

Porat, M. and Zeevi, Y. Y. (1989). Localized texture

processing in vision: Analysis and synthesis in the ga-

borian space. IEEE Transactions on Biomedical En-

gineering, 36(1):115–129.

Randen, T. and Husy, J. H. (1999). Filtering for texture

classiﬁcation: a comparative study. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

21(4):291–310.

Savari, S. A. (1997). Redundancy of the lempel-ziv-welch

code. In Proc. DCC, Salt Lake City, pages 191–200.

Shannon, C. E. (1948). A mathematical theory of commu-

nication. Bell Syst. Tech. J., 27:379–423.

Smith, J. R. and Chang, S.-F. (1997). Visually searching the

web for content. IEEE MultiMedia, 4(3):12–20.

Southard, T. E. and Southard, K. A. (1996). Detection of

simulated osteoporosis in maxillae using radiographic

texture analysis. IEEE Transactions on Biomedical

Engineering, 43(2):123–132.

Teahan, W. J. and Harper, D. J. (2001). Using compression-

based language models for text categorization. In

Workshop on Language Modeling and Information

Retrieval, pages 83–88.

Welch, T. A. (1984). A technique for high performance data

compression. IEEE Computer, 17(6):8–19.

Xu, K., Georgescu, B., Meer, P., and Comaniciu, D. (2000).

Performance analysis in content-based retrieval with

textures. In Proc. Int. Conf. Patt. Recog., volume 4,

pages 275–278.

Zibreira, C. (2000). Descric¸

ao e Procura de V

ıdeo

Baseadas na Forma. Master thesis, Instituto Superior

ecnico de Lisboa, Portugal.

Ziv, J. (1988). On classiﬁcation with empirically observed

statistics and universal data compression. IEEE Trans-

actions on Information Theory, 34(2):278–286.

Ziv, J. and Lempel, A. (1978). Compression of individ-

ual sequences via variable-rate coding. IEEE Trans-

actions on Information Theory, 24(5):530–536.

Ziv, J. and Merhav, N. (1993). A measure of relative en-

tropy between individual sequences with application

to universal classiﬁcation. IEEE Transactions on In-

formation Theory, 39(4):1270–1279.

VISAPP 2006 - IMAGE UNDERSTANDING