IMAGE RECOGNITION, ANALYSIS,

UNDERSTANDING, AND PROCESSING TECHNIQUES

A New Method for Coarse Classification of Textures

and Class Weight Estimation for Texture Retrieval1
K. L. Lee and L. H. Chen*
Department of Computer and Information Science, National Chiao Tung University
1001 Ta Hsueh Rd., Hsinchu, Taiwan 30050, Republic of China
e-mail:[email protected]
* To whom all correspondence should be addressed

Abstract—In this paper, a new texture classification method is provided for dividing texture images into three
classes: periodic, directional, and random. The method is based on the fact that for a directional texture image,
the magnitudes of its Fourier spectrum will concentrate on a certain direction; for periodic, on several direc-
tions; and for random, spread out over all directions. To use this fact, Fourier transform is, first, performed. Prin-
cipal component analysis is, then, conducted on the Fourier spectrum to get the ratio of two eigenvalues, which
will be used to measure the directionality of the texture image. If the texture image is not a directional one, for
getting better discriminative properties for separating periodic textures from random ones, Fourier transform is
applied to the Fourier spectrum to produce an enhanced Fourier spectrum. A discriminative measure based on
the variance of the radial wedge distribution is, then, calculated and applied to classify the texture image as peri-
odic or random one. Texture images from Brodatz album and Corel image database are used to demonstrate the
effectiveness of the proposed method. In addition, it is shown that the intermediate results of the proposed
method can be used to derive a weighting scheme used for texture retrieval. The proposed method can also be
used to implement the texture browsing descriptor of MPEG-7.

1
INTRODUCTION plexity. On the other hand, texture modeling based on
Wold decomposition has been proposed by Francos
Texture presents almost everywhere in natural and et al. [2–3]. According to Wold decomposition, a 2D
real world images and, therefore, has long been an homogeneous random field is decomposed into three
important research topic in image processing. Success- mutually orthogonal components: periodicity, direc-
ful applications of texture analysis methods have been tionality and randomness which are consistent with the
widely found in industrial, biomedical, and remote three most important dimensions of human texture per-
sensing areas. In addition, the recent emerging of mul- ception. If texture images can be coarsely classified
timedia and the availability of large image and video into these three categories, texture features can, then,
archives have made content-based information retrieval be chosen or designed specifically for each category. To
a very popular research topic. Texture is also deemed as be more specific, for periodic textures, the periodic fea-
one of the most important features when performing tures can be extracted by the methods specifically
content-based information retrieval. Various textural designed for periodic textures [5–6]; for random tex-
features have been adopted to fulfill these applications. tures, MRSAR model [7] is reported to have the best
Since there are a lot of variations among natural tex- performance and can be used for texture discrimination
tures, to achieve the best performance for texture anal- applications. The texture retrieval method proposed by
ysis or retrieval, different features should be chosen Liu and Picard [4] used this idea and demonstrate a bet-
according to the characteristics of texture images. ter performance. It provides a pre-classification step
Therefore, developing an effective method for prelimi- assigning weights to the two classes of texture images:
nary texture classification based on the textural charac- periodic and non-periodic. The weight of each class
teristics will greatly help the design of a texture classi- stands for the probability that the texture image belongs
fication system or a content-based texture retrieval sys- to the class. Our method also provides a weighting
tem. scheme for three classes of texture images: directional,
Rao and Lohse [1] conducted a texture study based periodic, and random. Thus, the proposed method pro-
on human perception, they conclude that the most vides a finer pre-classification than that in [4].
important dimensions of natural texture discrimination The properties of the Fourier spectrum of textures
are periodicity, directionality, granularity, and com- have been well studied [9–11], they can be summarized
1 This article was submitted by the authors in English. as follows: (1) for periodic textures, the Fourier spec-
trum consists of significant peaks scattered out regu-
larly at some directions; (2) for textures with strong
directionality, the directionality will be preserved in the
Received June 7, 2002 Fourier spectrum; (3) for random textures, the distribu-
Pattern Recognition and Image Analysis, Vol. 12, No. 4, 2002, pp. 400–410.
Original Text Copyright © 2002 by Pattern Recognition and Image Analysis. English Translation Copyright © 2002 by MAIK “Nauka /Interperiodica” (Russia).
 A NEW METHOD FOR COARSE CLASSIFICATION 401

tions of the responses of spectrum are not restricted to odic or random one. In the following, the proposed
certain directions. The proposed method is developed method will be described in detail. Besides, we will use
on the basis of these properties, it consists of two the intermediate results of the proposed method to pro-
phases: (1) directionality classification; (2) periodicity vide a weighting scheme for texture retrieval.
and randomness classification. In the phase of direc-
tionality classification, Fourier transform is, first, per-
formed on the texture image to obtain its Fourier spec- 2.1. Directionality Classification Phase
trum. Principal component analysis is, then, conducted
to get two eigenvalues. If the texture image contains Some examples of the properties of the Fourier
strong directionality, then the larger eigenvalue will be spectra of textures are shown in Fig. 2. Figure 2a shows
much greater than the smaller eigenvalue. Based on this a periodic texture (D1, from Brodatz album). As shown
phenomenon, the ratio of the larger eigenvalue to the in Fig. 2d, the spectral peaks are spread out regularly in
smaller eigenvalue is used to measure the directionality horizontal and vertical directions. The texture image
of the texture image. If the texture image is not classi- shown in Fig. 2b is a vertically directional texture
fied as a directional one, the periodicity and random- (D106). As shown in Fig. 2e, the spectral peaks form a
ness classification phase is entered. Fourier transform is horizontal line-like shape. Figure 2c shows a random
applied to the Fourier spectrum to obtain an enhanced texture (D54). Its Fourier spectrum is shown in Fig. 2f.
Fourier spectrum. The enhanced Fourier spectrum has The distribution of the spectral values in the Fourier
more discriminative properties in separating periodic spectrum is not restricted to any direction. In addition
textures from random ones than Fourier spectrum. For to the textural features described above, it is observed
periodic textures, those points with high magnitude that some noise exists on the Fourier spectra. To reduce
appear in some directions more clearly than those in the the noise, a Gaussian smoothing filter is applied to the
Fourier spectrum. Based on the properties, a discrimi- Fourier spectrum (see Figs. 2g–2i)).
native measure is, then, calculated to classify the tex-
ture image as a periodic or a random texture. Texture From the smoothed Fourier spectrum, we can see
images from Brodatz album [8] and Corel image data- that for directional textures, if we locate high spectral
base are used to demonstrate the effectiveness of the pixels and find their two principal components, the
proposed method. It is also shown that the intermediate direction of the eigenvector with larger eigenvalue will
results of the proposed method can be used to derive the be aligned with the line-like shape observed. In addi-
weights used for texture retrieval. tion, the larger eigenvalue will be much greater than the
smaller eigenvalue. Therefore, the ratio of the larger
The rest of the paper is organized as follows. In Sec- eigenvalue to the smaller eigenvalue can be used to
tion 2, the proposed method is described in detail and a measure the directionality of the texture image.
weighting scheme for texture retrieval is provided.
Experimental results and discussion are presented in To get the ratio, a thresholding method is, first,
Section 3. Finally, in Section 4, we give a conclusion. designed to extract the high spectral pixels from the
Fourier spectrum. Assume that the spectral values of
Fourier spectrum range from 0 to Lmax . Based on the
2. THE PROPOSED METHOD fact that the high spectral pixels contribute most of the
The system block of the proposed method is shown energy, we can locate these pixels through a threshold
in Fig. 1. Basically, the proposed method consists of value l with the aggregated energy from Lmax to l greater
two phases: (1) directionality classification and than t% of the total spectrum.
(2) periodicity and randomness classification. In direc- The high spectral pixel thresholding algorithm is
tionality classification phase, Fourier transform is, first, presented as follows.
performed on the texture image. To reduce noises, the
Fourier spectrum is smoothed. A thresholding method Algorithm 1. High spectral pixel thresholding
is, then, applied on the smoothed Fourier spectrum to
extract pixels with high spectral value. Principal com- Input. The smoothed Fourier spectrum F'(u, v), u =
ponent analysis is, then, conducted on the high spectral 1, …, N, v = 1, …, N of the texture image f(x, y) with
pixels and two eigenvalues are calculated. The ratio of 0 ≤ F'(u, v) ≤ Lmax .
the two eigenvalues is used to measure the directional- Output. The high spectral pixel set, H.
ity of the texture image. If the texture image is not clas-
sified as a directional one, the classification phase of Step 1. Evaluate the histogram h(i), i = 0, …, Lmax
periodicity and randomness is entered. Fourier trans- of F'.
form operation is, first, performed on the smoothed
Fourier spectrum to enhance the distinction between Step 2. Compute the energy E of F' as:
periodic and random textures. On the basis of the
N N
enhanced Fourier spectrum, a discriminative measure
based on the variance of the radial wedge distribution E = ∑ ∑ F' ( u, v ),
is, then, provided to classify the texture image as a peri- u = 1v = 1

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

402 LEE, CHEN

Input the
Texture Image

Perform
Fourier Transform and
Smooth the Fourier
Spectrum

Threshold the Smoothed

Fourier Spectrum

Perform Principal Calculate Enhanced

Component Fourier Spectrum
Analysis

Calculate Calculate
Directionality Discriminative
Measure Measure

Directional NO Periodic
Texture? Texture?

YES YES NO

Classify Classify Classify

the Texture the Texture the Texture
as Directional as Periodic as Random

Directionality Classification Periodicity and Randomness Classification Phase

Phase

Fig. 1. The system block of the proposed method.

Step 3. Let CEj be the cumulative energy of histo- We, then, calculate the principal components of H.
gram h(i) from j to Lmax and be evaluated as follows: To emphasize the importance of pixels with higher
spectral value in the calculation of principal compo-
L max
nent, we use the spectral values of pixels as weights
CE j = ∑ h ( i ). when calculating the principal components. First, the
co-variance matrix is calculated as
i= j

Determine the largest threshold l which satisfies the c uu c vu

following criterion: C = . (1)
c uv c vv
l = max{j|CEj /E > t1}.
∑
1
Step 4. Use l to threshold F'(u, v) as follows: where cuu = ----- ( u, v ) ∈ H F' (u, v)(u – u )2, cvv =
W
If F'(u, v) < l then (u, v) is not a high spectral pixel;
∑
1
----- ( u, v ) ∈ H F' (u, v)(v – v )2, cuv =
else assign (u, v) to H. W

∑
End if. 1
----- ( u, v ) ∈ H F' (u, v)(u – u )(v – v )2,cuv = cvu and
End of algorithm 1. W

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

 A NEW METHOD FOR COARSE CLASSIFICATION 403

(a) A periodic texture (b) A directional texture (c) A random texture

(D1). (D106). (D106).

(d) The Fourier spectrum (e) The Fourier spectrum (f) The Fourier spectrum
of (a). of (b). of (c).

(g) The smoothed Fourier (g) The smoothed Fourier (g) The smoothed Fourier
spectrum of (d). spectrum of (e). spectrum of (f).
Fig. 2. Some examples for the Fourier spectra of periodic, directional, and random textures.
(a) Periodic texture (D1).
(b) Directional texture (D106). (c) Random texture (D54).
(d) Fourier spectrum of (a). (e) Fourier spectrum of (b). (f) Fourier spectrum of(c).
(g) Smoothed Fourier spectrum of (d). (h) Smoothed Fourier spectrum of (e). (i) Smoothed Fourier spectrum of (f).

W= ∑ ( u, v ) ∈ H
F' (u, v), u = ∑( u, v ) ∈ H
F' (u, v)u, v = shown in Figs. 3a and 3b, respectively. We model the
distribution of EVs in each class by a Gaussian distribu-
----- ∑
1 tion. The directional class is denoted by wd and the non-
( u, v ) ∈ H
F' (u, v) v .
W directional class is denoted by wnd . The prior probabilities
Then the two eigenvalues of C are evaluated, and let are denoted by p(wd ) and p(wnd ), respectively. The condi-
them be λ1 and λ2 with λ1 > λ2 . The eigenvalue ratio tional probability density functions of EV’s are denoted by
EV= λ1/λ2 is, then, used to measure the directionality of p(EV|wd ) and p(EV|wnd ). And their mean and variance are
textures. estimated by the unbiased sample mean and sample vari-
The directionality classifier is designed using a ance of the EV’s of the training set.
Bayesian approach. To design the classifier, some sam- Given a texture, the posterior probability of wd is
ple images from Brodatz album are used as training set. computed as
Then, EV ratio is computed for each image in the train-
p ( EV , w d ) p ( EV w d ) p ( w d )
ing set. The distributions of EV ratios for the non-direc- p ( w d EV ) = ------------------------- = ---------------------------------------
-. (2)
tional and directional images in the training set are p ( EV ) p ( EV )

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

404 LEE, CHEN

(b) features from the Fourier spectrum. By applying Fou-

(a) rier transform to Fig. 4e, the spectral peaks appear more
count

prominent on the enhanced Fourier spectrum shown in

Fig. 4f. Figure 4g shows a random texture. Its Fourier
spectrum is shown in Fig. 4h. The spectral responses in
the Fourier spectrum are not periodic and do not con-
EV EV centrate on certain frequencies but scatter around the
frequency plane. That is why pixels of a random texture
Fig. 3. The distributions of the EV’s of the images in the are not of certain periods, thus, they will not contribute
training set.
(a) EV distribution of non-directional class. (b) EV distribu-
to certain frequencies and form periodic spectral peaks.
tion of directional class. This also makes the spectral responses in the enhanced
Fourier spectrum (see Fig. 4i) spread over all direc-
tions. According to the above-mentioned properties, the
We can then design the directionality classifier as fol- enhanced Fourier spectrum is adopted to discriminate
lows: periodic textures from random ones.
If p(wd|EV)p(wnd|EV), classify the texture as direc- Based on enhanced Fourier spectrum, Radial Wedge
tional; otherwise classify the texture as non-directional. Distribution Variance (RWDV) will be defined and used
Through Eq. (2), the inequality equation in the if to design a discriminative measure. In the following,
condition of the classifier can be replaced by we will explain RWDV in details.
p(EV|wd )p(wd ) ≥ p(EV|wnd)p(wnd). If the texture image Given an enhanced Fourier spectrum, E(u, v), its
is classified as non-directional, the classification pro- radial wedge distribution is first calculated. Figure 5
cess enters into periodicity and randomness classifica- shows the radial wedges used in the proposed method.
tion phase. Let the radial wedges be denoted by RWi , i = 1, …, m,
m = 360/∆θ, where ∆θ is the size of each wedge. For
2.2. Periodicity and Randomness Classification Phase each E(u, v), we accumulate it to RWi if it satisfies

In this phase, each input texture image is assumed to

θ i ≤ tan  ---- < θ i + 1 .
–1 u
be non-directional, we will classify it as periodic or ran- v
dom. As mentioned previously, the properties of the
Fourier spectra of textures can be further enhanced by
performing Fourier transform on the Fourier spectrum The energy of each wedge is, then, normalized by
again. The obtained Fourier spectrum is called an the total energy of all wedges. Let ERWi , i = 1, …, m
enhanced Fourier spectrum in this paper. The textural denote the normalized energy of radial wedges RWi ,
features of the enhanced Fourier spectrum are more then the RWDV is defined as:
prominent than those of the original Fourier spectrum.
m
Two examples are shown in Fig. 4. Figure 4a shows a
∑ ( ERW
1 2
periodic texture from D1. Figure 4b shows the Fourier RWDV = ---- i – ERW ) , (3)
spectrum of Fig. 4a, the spectral peaks, each of which m
i=1
comes from the contribution of those pixels with the
same period in the original image, spread out regularly
∑
1
along certain directions. This property is enhanced in where ERW = ---- m
i=1
( ERW i ) .
m
the enhanced Fourier spectrum, because those peaks in
the Fourier spectrum are periodic, through applying For periodic textures, as the spectral peaks spread
Fourier transform again, peaks with the same period out regularly along certain directions while the spectral
will contribute to the same frequency, thus, the peaks in values of random textures appear in all directions, the
the enhanced Fourier spectrum will be more prominent. variance of all radial wedge energies of a periodic tex-
On the other hand, regarding the remaining pixels (not ture will be larger than that of a random texture. There-
peaks) in the Fourier spectrum, since they are not peri- fore, RWDV can be used to separate periodic textures
odic, after the second applying of the Fourier trans- from random ones.
form, they do not contribute to the same frequency.
This phenomenon also relatively enhances those peaks. The classification follows the same Bayesian
Figure 4c shows the enhanced Fourier spectrum of approach proposed in Section 2.1. RWDV is computed
Fig. 4a. Figure 4d is the image obtained by adding 30% for each image in the training set. The distributions of
of Gaussian noise in Fig. 4a. Figure 4e is the smoothed RWDV’s for the random and periodic images in the
Fourier spectrum of Fig. 4d, the frequencies to which training set are shown in Figs. 6a and 6b, respectively.
the periodic patterns contribute, are mixed with noise in Each distribution of RWDV’s is approximated by a
the Fourier spectrum. This is the reason of heavy non- Gaussian distribution. The periodic class is denoted by
periodic noise. Thus, it is difficult to extract textural wp and the random class is denoted by wr . Then, the tex-

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

 A NEW METHOD FOR COARSE CLASSIFICATION 405

(a) A periodic texture (b) The smoothed Fourier (c) The enhanced Fourier
(D1). spectrum of (a). spectrum of (a).

(d) The image after adding (e) The smoothed Fourier (f) The enhanced Fourier
30% of noise on (a). spectrum of (d) spectrum of (d).

(g) A random texture (g) The smoothed Fourier (g) The enhanced Fourier
(D54). spectrum of (g). spectrum of (g).
Fig. 4. Two examples of the enhanced Fourier spectra for periodic and random textures.
(a) Periodic texture (D1). (b) The smoothed Fourier spectrum of (a). (c) The enhanced Fourier spectrum of (a). (d) Image (a) after
adding 30% of noise. (e) The smoothed Fourier spectrum of (d). (f) The enhanced Fourier spectrum of (d). (g) Random texture
(D54). (h) The smoothed Fourier spectrum of (g). (i) The enhanced Fourier spectrum of (g).

ture periodicity and randomness classifier is designed 2.3. Weighting Scheme for Texture Retrieval
as follows: First, we will briefly describe the texture retrieval
If p(RWDV|wp )p(wp ) ≥ p(RWDV|wr )p(wr ), then system proposed in [4]. Given a query image, the sys-
tem uses a pre-classification method to evaluate the
classify the texture as periodic; otherwise, classify the probabilities (Pi , i = 1, 2) of the query image belonging
texture as random. to each of the two classes: periodic and non-periodic
By processing the texture image through the two textures. Next, the periodic and non-periodic features
phases of the proposed method, we can successfully are extracted. By using the ith kind of feature for
retrieval, the ordering vector Oi for all images in data-
classify the texture image as directional, periodic, or
base is determined. The joint ordering vector O for all
random. The intermediate results of the proposed images is finally obtained by summing the results of
method can be used to develop a weighting scheme for multiplying the ordering vectors O1 and O2 by the
texture retrieval. This idea has been used in the texture weights P1 and P2 . That is, O = P1O1 + P2O2 . The sys-
retrieval method proposed by Liu and Picard [4], the tem can get better performance without using the
details will be explained in the following section, weights.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

406 LEE, CHEN

probabilities p(wp|RWDV) and p(wr |RWDV) standing

RW3 for the posterior probability of wp and wr , respectively,
RW4 RW2
will be computed. Let wtd , wtp , and wtr denote the prob-
abilities of the texture image belonging to the classes of
RW5 RW1 directional, periodic, and random, respectively. Then,
the vector (wtd , wtp , wtr) can be evaluated as
RW6 RW10
wtd = p(wd|EV),
RW7 RW9 wtp = p(wnd|EV) ∗ p(wp |RWDV),
RW8
wtr = 1 – wtd – wtp .
Fig. 5. The radial wedges used in the proposed method. Since the proposed method provides a finer pre-
classification than that in [4], this should help to
improve the performance of a texture retrieval system.
The reason is that the final performance of a texture
(a) (b) retrieval system heavily depends on the textural fea-
tures related to the texture classes. The system flow of
a texture retrieval system using the weight vector (wtd ,
count

wtp , wtr) is shown in Fig. 7.

3. EXPERIMENTAL RESULTS
Fig. 6. The distributions of the RWDV’s of the images in tra- Texture images of Brodatz album and Corel Gallery
ming set. image database were used to test the proposed method.
(a) The RWDV distribution of the random class. (b) The To build up the Brodatz album database, eight patches
RWDV distribution of the periodic class.
for each of the 112 textures in Brodatz album were
scanned and 896 texture images were obtained for
As illustrated in the previous sections, the proposed experiments. Four out of the eight patches of each tex-
method provides a coarse classification of texture ture were used as training set to obtain the empirical
images into three classes: periodicity, directionality, values for parameters used, while the remaining images
were used as a testing set. To further validate the perfor-
and randomness. In the following, a weighting scheme
mance of the proposed method, 1896 natural color tex-
associated with this coarse classification will be pro- ture images from Corel Gallery image database were
vided. For a texture image, in the directionality classi- selected and used as a testing set as well, including
fication phase, two probabilities of the texture image, abstract textures, bark textures, creative textures, food
p(wd|EV) and p(wnd|EV) standing for the posterior prob- textures, light textures, etc. Some examples of Corel
ability of wd and wnd, respectively, will be calculated. If Gallery database are shown in Fig. 8. We will report the
the texture image is non-directional, in the periodicity experimental results of Brodatz database and Corel
and randomness classification phase, two additional database, respectively.

Extract the wtd

Directional Directional
Features Ordering

Calculate the wtp

The Query Texture
Weight Vector Extract the Joint Retrieved Texture
Image Periodic
Using the Periodic Features Ordering Images
Ordering
Proposed
Method

Extract the wtr

Random
Random Features Ordering

Fig. 7. The revised system flow of the texture retrieval system presented in [4] using the proposed method.

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

 A NEW METHOD FOR COARSE CLASSIFICATION 407

(a) An abstract texture. (b) A bark texture. (c) A creative texture.

(d) A food texture. (e) A light texture. (f) A flower texture.

(g) A cloud texture. (h) A candy texture. (i) A brick texture.

Fig. 8. Some textures of Corel Gallery image database used in the experiments.
(a) Abstract texture. (b) Bark texture. (c) Creative texture. (d) Food texture. (e) Light texture. (f) Flower texture. (g) Cloud texture.
(h) Candy texture. (i) Brick texture.

To classify the texture images in the testing set into tively. Figure 9 shows some examples of correctly clas-
directional, periodic, or random, we, first, classify the sified directional, periodic, and random texture images.
textures into directional or non-directional ones (step 1). Four images are shown for each category, the first two
Those classified as non-directional are then further images are from Brodatz database and the latter two
classified into periodic or random (step 2). By inspect- images are from Corel database.
ing the images in the testing set, the classification rate Tables 1 and 2 both show that the classification rates
of each step as well as the estimated and actual classi- of step 1 (99.3% and 99.8%) are quite high, this dem-
fication rates are reported. The estimated classification onstrates the effectiveness of the proposed method in
rate is obtained by multiplying the classification rates discriminating directional textures from non-direc-
of both steps. The actual classification rate is the total tional textures. Some of the misclassified textures in
number of correctly classified images via both steps step 1 are shown in Fig. 10. Figure 10a is a periodic tex-
divided by the total number of images in the testing set. ture image classified as directional. It can be noticed
The classification result of Brodatz database and Corel that although both vertical and horizontal lines are
database are summarized in Tables 1 and 2, respec- present in the image, the horizontal lines are not signif-

Table 1. The performance for the classification of Brodatz Table 2. The performance for the classification of Corel
database database
Step 1 Step 2 Step 1 Step 2
Step Classification Rate 99.3% 96.1% Step Classification Rate 99.8% 98.8%
Estimated Classification Rate 95.4% Estimated Classification Rate 98.6%
Actual Classification Rate 95.5% Actual Classification Rate 98.8%

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

408 LEE, CHEN

(a) Some correctly classified directional textures.

(b) Some correctly classified periodic textures.

(c) Some correctly classified random textures.

Fig. 9. Some correctly classified texture images from Brodatz database and Corel database.
(a) Some correctly classified directional textures. (b) Some correctly classified periodic textures. (c) Some correctly classified ran-
dom textures.

(a) (b) (c) (d)

Fig. 10. Some misclassified textures and their Fourier spectra of step 1.

icant enough. Thus, the high spectral pixels of its Fou- be more accurate in classifying textures, we can intro-
rier spectrum (Fig. 10b) form a horizontal line-like duce additional classes named multi-directional to cope
region, making it misclassified. Figure 10c shows a with diversity of natural textures. Some examples of
directional texture classified as non-directional. Since multi-directional textures from Corel database are
there are only three classes used, we consider Fig. 10e shown in Fig. 11.
as a directional texture. However, there are actually Similarly, some of the misclassified textures of step
groups of straws arranged in four different directions. 2 are shown in Fig. 12. Figure 12a shows a wave texture
Therefore, its Fourier spectrum shown in Fig. 10d also and is classified as periodic; however, we consider it as
presents four lines distributed in different directions. a random texture. It is observed that, in addition to most
This makes Fig. 10c be classified as non-directional in of the homogeneous areas, there are directional wave-
step 1, and in step 2, Fig. 10c will be further classified like patterns present in the image. The classification
as periodic. This classification error is caused by a error is due to the fact that, although Fig. 12a is not
small number of classes used. In fact, Fig. 10c is neither close to any class, it has to be classified to one of the
directional nor periodic, it is multi-directional. Thus, to three classes used. Fortunately, the class probabilities

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

 A NEW METHOD FOR COARSE CLASSIFICATION 409

(a) (b) (c) (d)

Fig. 11. Some examples of multidirectional textures from Corel database.

(a) (b) (c)

Fig. 12. Some misclassified textures and their enhanced Fourier spectra of step 2.

(a) [0.98] (b) [0.73] (c) [0.15] (d) [0.11]

Fig. 13. Four textures with clear primitives but different displacements.

of Fig. 12a for periodic and random are 0.53 and 0.47, Section 2.3 is consistent with human perception. In
respectively. Since these two probabilities are quite addition, a descriptor for measuring the regularity of
close, Fig. 12a can be considered as an ambiguous tex- textures has been specified in the texture browsing
ture. In practice, all images of this kind can be consid- descriptor of MPEG-7 [12]. The calculated periodic
ered as ambiguous. Figure 12b shows a periodic texture weight can also be used to implement this regularity
misclassified as random. Although Fig. 12b is per- descriptor.
ceived as havinge clear directionality and some period-
icity, some local variations are present and distort the
directionality and periodicity. Thus, the high spectral 4. CONCLUSION
pixels of its enhanced Fourier spectrum (Fig. 12c) do In this paper, we propose a new method for coarse
not concentrate at certain directions. This makes Fig. classification of textures and a weighting scheme for
12b be classified as random. texture retrieval. The eigenvalue ratio obtained by per-
Figure 13 shows four textures from Corel database. forming principal component analysis on the Fourier
They are all textures with clearly defined texture prim- spectrum of the texture image is, first, used to deter-
itives but different in the regularity of displacement. mine the directionality measure. If the texture image is
The regularity of their displacements is decreasing not a directional one, Fourier transform is applied to the
from Fig. 13a to Fig. 13d. The corresponding class Fourier spectrum image to produce an enhanced Fou-
probability for periodic is listed under each figure. rier spectrum. A discriminative measure based on the
These values are also in decreasing order. This reveals variance of the radial wedge distribution is, then, calcu-
that the class weights calculation scheme proposed in lated and applied to classify the texture image as peri-

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002

410 LEE, CHEN

odic or random one. The proposed method provides a gressive Models, Pattern Recognition, 1992, vol. 25,
coarse classification of textures based on the three most no. 2, pp. 173–188.
important dimensions of human texture perception, i.e.,
8. Brodatz, P., Textures–A photographic for Artists and
periodicity, directionality, and randomness. The pro- Designers, New York: Dover, 1966.
posed method can be used to implement the texture
browsing descriptor of MPEG-7. Furthermore, for dif- 9. Levine, M.D., Vision in Man and Machine, New York:
ferent classes, system designers can design their own McGraw Hill, 1985.
texture features and use the immediate results, class
probabilities, of the proposed method as weights for 10. Chen, C.H., A Study of Texture Classification Using
classes to perform texture retrieval. In the future, we Spectral Features, Proc. 6th Int. Conf. on Pattern Recog-
nition, Munich, 1982, pp. 1074–1077.
will develop texture classification methods to cope with
multi-directional textures in the Corel database. 11. Connors, R.W. and Harlow, C.A., A Theoretical Com-
parison of Texture Algorithms, IEEE Transactions on
Pattern Analysis and Machine Intelligence, 1980, vol. 2,
ACKNOWLEDGMENTS no. 3, pp. 204–222.
This research was supported in part by the National 12. Text of ISO/IEC 15938-3 Multimedia Content Descrip-
Science Council of Republic of China, contract no. tion Interface–Part 3: Visual. Final Committee Draft,
NSC-87-2213-E-009-060.) ISO/IEC/JTC1/SC29/WG11, Doc. N4062, 2001.

REFERENCES
Ling-Hwei Chen. Born in Taiwan, Republic of China, in
1. Rao, A.R. and Lohse, G.L., Toward a Texture Naming 1954. Received her BSci Degree in Mathematics, her Msci
System: Identifying Relevant Dimensions of Texture, Degree in Applied Mathematics from National Tsing Hua
Proc. IEEE Conf. Visualization, 1993, pp. 220–227. University, Hsinchu, Taiwan in 1975 and 1977, respectively,
2. Francos, J.M., Zvi Meiri, A., and Porat, B., A Unified and her PhD Degree in Computer Engineering from National
Texture Model Based on a 2D Wold Like Decomposi- Chiao Tung University in 1987. In 1977-1979, was a
tion, IEEE Trans. on Signal Processing, 1993, pp. 2665- Research assistant in the Chung-Shan Institute of Science
2678. and Technology, Taoyan, Taiwan. In 1979-1981, was a
Research associate in the Electronic Research and Service
3. Sriram, R., Francos, J.M., and Pearlman, W.A., Texture Organization, Industry Technology Research Institute, Hsin-
Coding Using a Wold Decomposition Model, Proc. chu, Taiwan. In 1981-198, worked as an Engineer in the Insti-
ICPR, 1994, vol. 3, pp. 35–39. tute of Information Industry, Taipei, Taiwan. Professor of the
4. Liu, F. and Picard, R., Periodicity, Directionality, and Department of Computer and Information Science at the
Randomness: Wold Features for Image Modeling and National Chiao Tung University. Her current research inter-
Retrieval, IEEE Trans. on Pattern Analysis and Machine ests include image processing, pattern recognition, document
Intelligence, 1996, vol. 18, no. 7, pp. 722–733. processing, image compression, and image cryptography.
5. Kim, H.B. and Park, R.H., Extracting Spatial Arrange-
ment Of Spectral Textures Using Projection Information,
Pattern Recognition, 1992, vol. 25, no. 3, pp. 237–247. Kuen-Long Lee. Received both his BSci and MSci
6. Lin, H.C., Wang, L.L., and Yang, S.N., Extracting Peri- Degrees in Computer and Information Science from National
odicity Of A Regular Texture Based On Autocorrelation Chiao Tung University, Taiwan, in 1988 and 1990, respec-
Functions, Pattern Recognition Letters, 1997, vol. 18, tively. He is currently a PhD student in the Department of
pp. 433–443. Computer and Information Science at National Chiao Tung
7. Mao, J. and Jain, A.K., Texture Classification and Seg- University. His research interests include image processing,
mentation Using Multiresolution Simultaneous Autore- pattern recognition, texture analiysis, and image retrieval.

SPELL: directionality, thresholding, co-variance, misclassified, wave-like, havinge, analiysis, lighttexture

PATTERN RECOGNITION AND IMAGE ANALYSIS Vol. 12 No. 4 2002