Pattern Recognition 32 (1999) 477—486

Unsupervised texture segmentation using feature distributions

Timo Ojala, Matti Pietikäinen
Machine Vision and Media Processing Group, Infotech Oulu, University of Oulu, FIN-90570 Oulu, Finland
Received 19 December 1997; in revised form 24 February 1998

Abstract

This paper presents an unsupervised texture segmentation method, which uses distributions of local binary patterns
and pattern contrasts for measuring the similarity of adjacent image regions during the segmentation process. Non-
parametric log-likelihood test, the G statistic, is engaged as a pseudo-metric for comparing feature distributions.
A region-based algorithm is developed for coarse image segmentation and a pixelwise classification scheme for improving
localization of region boundaries. The performance of the method is evaluated with various types of test images.  1999
Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.

Keywords: Texture segmentation; Feature distribution; G statistic; Spatial operator; Local binary pattern; Contrast

1. Introduction include approaches based on split-and-merge methods

[5], pyramid node linking [6], selective feature smooth-
Segmentation of an image into differently textured ing with clustering [7], and a quadtree method combin-
regions is a difficult problem. Usually, one does not know ing statistical and spatial information [8]. Examples of
a priori what types of textures exist in an image, how more recent approaches are methods based on local
many textures there are, and what regions have which linear transforms and multiresolution feature extraction
textures [1]. In order to distinguish reliably between two [9], feature smoothing and probabilistic relaxation [10],
textures relatively large samples of them must be exam- autoregressive models [11,12], Markov random field
ined, i.e. relatively large blocks of the image. But a large models [13—16], multichannel filtering [17—19], neural
block is unlikely to be entirely contained in a homogene- network-based generalization of the multichannel ap-
ously textured region and it becomes difficult to correctly proach [20], wavelets [21,22], fractal dimension [23],
determine the boundaries between regions. and hidden Markov models [24]. A method for unsuper-
Many different approaches to image and texture seg- vised segmentation of color textures using Markov ran-
mentation have been proposed [2—4]. Segmentation dom fields and a split-and-merge type algorithm was
methods are usually classified as region-based, bound- proposed by Panjwani and Healey [25].
ary-based or as a hybrid of the two. The segmentation Some of the existing methods perform reasonably well
can be supervised or unsupervised. In unsupervised seg- for a small set of fine-grained texture mosaics, but they
mentation no a priori information about the textures usually need some prior knowledge of the image contents
present in the image is available. This makes it is a very to achieve satisfactory results, like the number of textures
challenging research problem in which only limited suc- or regions. The choice of proper parameters for different
cess has been achieved so far. Early methods proposed types of images may also be difficult and the methods
for unsupervised region-based texture segmentation typically perform poorly for natural images containing

0031-3203/99/$ — See front matter  1999 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.
PII: S 0 0 3 1 - 3 2 0 3 ( 9 8 ) 0 0 0 3 8 - 7
478 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486

nonuniform textures. Unsupervised segmentation of im- describes the spatial structure of the local texture, but it
ages containing texture primitives at very different scales does not address the contrast of the texture. For this
may even be unrealistic, because it is hard to discriminate purpose we combine LBP with a simple contrast measure
small image regions from large texture primitives without C, which is the difference between the average gray-level
any prior knowledge. of those pixels which have value 1 and those which have
The choice of highly discriminating texture features is value 0 (Fig. 1b).
the most important factor for a success in texture seg- The LBP/C distribution is approximated by a discrete
mentation, but this has been neglected in most earlier two-dimensional histogram of size 256;b, where b is the
approaches. The features should easily discriminate vari- number of bins for C. Choosing b is a trade-off between
ous types of textures and the window size used for com- the discriminative power and the stability of the texture
puting textural features should be small enough to be transform. If b is too small, the histogram will lack
useful for small image regions and to provide small error resolution and feature C will add very little dis-
rates at region boundaries. criminative information to the process. However, since
Our recent studies show that excellent texture dis- the image region contains a finite number of pixels, it
crimination can be obtained with local texture operators does not make sense to go to the other extreme, for then
and nonparametric statistical discrimination of sample the histogram becomes sparse and unstable. Based on the
and prototype distributions. Texture classification re- results of our past texture classification experiments with
sults obtained by using distributions of local binary the LBP/C transform, we chose to use 8 bins, although
patterns (LBP) or gray-scale differences have been we expect to achieve comparable results with 4 or 16 bins
better than those obtained with the existing methods as well. See Ref. [27] for a detailed description of the
[26—29]. Our method can be easily generalized to utilize mapping from the continuous C space to the discrete bin
multiple texture features, multiscale information, color index.
features and combinations of multiple features using A log-likelihood-ratio, the G statistic [30], is used as
the new multichannel approach to texture descrip- a pseudo-metric for comparing LBP/C distributions. The
tion [29]. value of the G statistic indicates the probability that the
This paper presents an efficient method for unsuper- two sample distributions come from the same popula-
vised texture segmentation based on texture description tion: the higher the value, the lower the probability that
with feature distributions. A region-based algorithm is the two samples are from the same population. We mea-
developed for coarse image segmentation and a pixelwise sured the similarity of two histograms with a two-way
classification scheme for improving the localization of test of interaction or heterogeneity:
region boundaries.

  


L L L
G"2 f log f ! f log f
G G G G
2. Texture description QK G QK G G




L
The texture contents of an image region are character- ! f log f
G G
ized by the joint distribution of local binary pattern G QK QK
(LBP) and contrast (C) features [27]. The original 3;3




neighborhood (Fig. 1a) is thresholded by the value of the L L
# f log f (1)
center pixel. The values of the pixels in the thresholded G G
QK GK QK G
neighborhood (Fig. 1b) are multiplied by the binomial
weights given to the corresponding pixels (Fig. 1c) and where s, m are the two sample histograms, n is the
obtained values (Fig. 1d) are summed for the LBP num- number of bins and f is the frequency at bin i. The more
G
ber (169) of this texture unit. By definition LBP is invari- alike the histograms s and m are, the smaller is the value
ant to any monotonic gray-scale transformation. LBP of G.

Fig. 1. Computation of local binary pattern (LBP) and contrast measure C.

 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486 479

3. Segmentation algorithm segments containing several textures are assumed to be

uniform. Threshold X was experimentally set to value
The segmentation method consists of three phases: 1.2. We computed G and G [Eq. (2)] for numerous
 
hierarchical splitting, agglomerative merging and pixel- blocks with varying texture contents, and a 20% differ-
wise classification. First, hierarchical splitting is used to ence generally indicated a deviation in the local texture.
divide the image into regions of roughly uniform texture. To begin with, we divide the image into rectangular
Then, agglomerative merging procedure merges similar blocks of size S . If we applied the uniformity test on

adjacent regions until a stopping criterion is met. At this arbitrarily large image segments, we could fail to detect
point we have obtained rough estimates of the different small texture patches and end up treating regions con-
textured regions present in the image and complete the taining several textures as uniform. The next step is to use
analysis by a pixelwise classification to improve the local- the uniformity test. If a block does not satisfy the test, it is
ization. Fig. 2 illustrates the progress of the segmentation divided into four subblocks. This procedure is repeated
algorithm on a 512;512 mosaic containing five different recursively on each subblock until a predetermined min-
Brodatz [31] textures. imum block size S is reached. It is necessary to set

a minimum limit for the block size, for the block has to
3.1. Hierarchical splitting contain a sufficient number of pixels for the LBP/C
histogram to be reliable. Since it is fundamental to make
A necessary prerequisite for the agglomerative merging reliable merges in the early stages of the merging process,
to be successful is that the individual image regions are we decided to use the relatively large value of 16 for S .

uniform in texture. For this purpose we apply the hier- Comparable results were obtained with value 8, whereas
archical splitting algorithm, which recursively splits the histograms of 4;4 blocks turned out to be too noisy in
original image into square blocks of varying size. The some cases. The choice of S is less crucial, and we

decision whether a block is split to four subblocks is chose to use 64. Fig. 2b illustrates the result of the
based on a uniformity test. We measure the six pairwise hierarchical splitting algorithm. As expected, the splitting
G distances between the LBP/C histograms of the four goes deepest around the texture boundaries.
subblocks. If we denote the largest of the six G values by Note that the hierarchical splitting phase is not man-
G and the smallest by G , the block is found to be datory, but we could skip it by dividing the input image
 
nonuniform and is thus split further into four subblocks, directly to blocks of size S and the successive agglom-

if a measure of relative dissimilarity within region is erative merging phase would still succeed. This is parti-
greater than a threshold cularly true for easier problems of homogeneous and
clearly distinct textures. However, our experiments have
G shown that finding larger areas of uniform texture with
R" 'X. (2)
G the hierarchical splitting method improves the conver-

gence of the agglomerative merging algorithm.
Regarding the proper choice of X, one should rather
choose a too small value for X instead of a too large one. It
is better to split too much than too little, for the following 3.2. Agglomerative merging
agglomerative merging procedure is able to correct errors,
where an uniform block of a single texture has been Once the image has been split into blocks of roughly
needlessly split. But error recovery is not possible, if uniform texture, we apply an agglomerative merging

Fig. 2. Texture mosaic C1; the main sequence of the proposed segmentation algorithm.
480 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486

procedure, which merges similar adjacent regions until

a stopping criterion is satisfied. At a particular stage of
the merging, we merge that pair of adjacent segments,
which has the smallest merger importance (MI) value.
MI is defined as

MI"p;G, (3)

where p is the number of pixels in the smaller of the two

regions and G is the distance measure defined in Eq. (1).
In other words, at each step the procedure chooses that
merger of all possible mergers, which introduces the
smallest change in the segmented image. Once the pair of
adjacent segments with the smallest MI value has been
found, the regions are merged and the two respective
LBP/C histograms are summed to be the histogram of
the new image region. Before moving to the next merger
we compute the G distances between the new region and
all adjacent regions to it. Merging is allowed to proceed
until the stopping rule
Fig. 3. Plot of MIR.
MI
MIR"  '½ (4)
MI
 small value of MIR underline the easiness with which

triggers. Merging is halted if MIR, the ratio of MI , the rough estimate of the texture regions is obtained for

merger importance for the current best merge, and mosaic C1. Note that the segmentation error of 1.4%
MI , the largest merger importance of all preceding after the agglomerative clustering phase (ERR ) is some-

mergers, exceeds a preset threshold ½. In theory, it is what biased in this problem, for the horizontal and verti-
possible that the very first merges have a zero MI value cal texture boundaries are accidentally aligned with the
(i.e. there are adjacent regions with identical LBP/C initial blocks.
histograms), which would lead to a premature termina-
tion of the agglomerative merging phase. To prevent this 3.3. Pixelwise classification
the stopping rule is not evaluated for the first 10% of all
possible merges. To improve the localization of the boundaries a simple
The value of threshold ½ was determined experi- pixelwise classification algorithm is used. If the hierarchi-
mentally. We applied the algorithm to numerous texture cal splitting and agglomerative merging phases have suc-
images and examined the values of MIR during the ceeded, we have obtained quite reliable estimates of the
merging process. The conclusion was that for well-de- different textured regions present in the image. Treating
fined homogeneous textures MIR values up to 1.5 or 1.6 the LBP/C histograms of the image segments as our
were still acceptable, while values over 2.0 generally were texture models we switch into a texture classification
due to a perceivable difference in the texture contents of mode. If an image pixel is on the boundary of at least two
the adjacent regions of the current best merge. Based on distinct textures (i.e. the pixel is 4-connected to at least
this observation we chose value 2.0 for ½. In more gen- one pixel with a different label), we place a discrete disc
eral terms, threshold ½ can be interpreted as the scale of with radius r on the pixel and compute the LBP/C
texture differences we want to discriminate, and thus the histogram over the disc. We compute the G distances
value of ½ may be a very subjective decision. This is between the histogram of the disc and the models of those
particularly the case with outdoor scenes and irregular regions, which are 4-connected to the pixel in question.
textures, where the number of distinct regions in the We relabel the pixel, if the label of the nearest model is
segmentation result strongly correlates with the thre- different from the current label of the pixel and there is at
shold. least one 4-connected adjacent pixel with the tentative
Fig. 2c shows the result of the agglomerative merging new label. The latter condition improves smooth adap-
phase after 174 merges. The MIR of the 175th merge tion of texture boundaries and decreases the probability
(MIR ) is 9.5 and the merging is halted. The highest of small holes occurring inside the regions. If the pixel is

MIR value up to that point (MIR ) had been 1.2 (Fig. 3). relabeled, i.e. it is moved from an image segment to the

The relationship between MIR , MIR and threshold adjacent segment, we update the corresponding texture
 
½ reflects the reliability of the result of the agglomerative models accordingly, hence the texture models become
merging phase. The very large value of MIR and very more accurate during the process. Only those pixels at

 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486 481

which the disc is entirely inside the image are examined,

hence the final segmentation result will contain a border
of r pixels wide.
In the next scan over the image we only check the
neighborhoods of those pixels, which were relabeled in
the previous sweep. The process of pixelwise classifica-
tion continues until no pixels are relabeled or maximum
number of sweeps is reached. This is set to be two times
S , based on the reasoning that the boundary estimate

of the agglomerative merging phase can be at most this
far away from the ‘‘true’’ texture boundary. Setting an
upper limit for the number of iterations ensures that the
process will not wander around endlessly, if the disc is
not able to capture enough information of the local
texture to be stable. According to our experiments the
algorithm generally converges quickly with homogene-
ous textures, whereas with locally stochastic natural
scenes maximum number of sweeps may be consumed.
We did not apply any postprocessing method to improve
Fig. 4. Final segmentation error as a function of the radius r of
the final segmentation result, e.g. by smoothing the tex-
the disc. The solid line corresponds to mosaic C5 and the
ture boundaries or removing small regions as many exist-
dashed line corresponds to mosaic C6.
ing algorithms do.
The relationship between the radius r of the disc and
the final segmentation result is obvious. A very small disc Table 1
is unstable, producing ragged texture boundaries and Parameters describing segmentation of mosaics C1 to C5
holes inside regions, whereas a very large disc is stable
and produces smooth boundaries, but may fail in locat- Mosaic MIR MIR ERR (%) ERR (%) Sweeps
  
ing the boundaries accurately. We used the final segmen-
tation error as a guide-line in choosing the value of r. C1 9.5 1.2 1.4 1.7 16
Mosaics C4 (Fig. 7) and C5 (Fig. 8) were processed with C2 5.2 1.6 4.2 1.2 23
the pixelwise classification algorithm, with r ranging C3 8.1 1.2 4.6 1.9 13
C4 4.1 1.3 2.0 1.4 9
from 1 to 20. The segmentation errors are plotted as
C5 2.8 1.2 7.8 2.1 24
a function of r in Fig. 4. As expected, the error first
decreases with increasing disc size, reaches the minimum
and then slowly increases as the disc becomes too large to
locate the boundaries accurately. Based on this result we scribing segmentation of texture mosaics are given in
chose value 11 for radius r. Table 1.
Fig. 2d shows the final segmentation result after the Mosaic C2 (Fig. 5a) is a 512;512 image containing
pixelwise classification phase, where 16 sweeps were four textures made by a GMRF process and a circle of
needed. The final segmentation error (ERR ), computed painted surface in the middle [32]. The more difficult

over the area processed by the disc which excludes the nature of this problem shows in the values of MIR

border of r pixels, is 1.7%. (5.2) and MIR (1.6), which are clearly closer to thre-

shold ½ than what was the case with mosaic C1. Never-
theless, the rough segmentation result (Fig. 5b) with
4. Experimental results segmentation error of 4.2% is quite decent. The final
segmentation result (Fig. 5c) after 23 sweeps with seg-
The segmentation results for four additional texture mentation error of 1.2% is excellent.
mosaics and two natural scenes are presented. The same Mosaic C3 (Fig. 6a) is a 512;512 image with a back-
set of parameter values was used for all texture mosaics ground made by a GMRF process and four distinct
to demonstrate the robustness of the approach: b"8, regions; the square and the circle are painted surfaces
S "64, S "16, X"1.2, ½"2.0, and r"11. In with different surface roughnesses and the ellipse and the
 
each case we provide the original image, the rough triangle are made by a fractal process [32]. As we can see
segmentation result after the agglomerative merging from the values of MIR (8.1) and MIR (1.2), the
 
phase and the final segmentation result after the pixel- rough estimate (Fig. 6b) of the texture regions is obtained
wise classification phase. The segmentation results are relatively easily. The final segmentation result (Fig. 6c)
superpositioned on the original image. Parameters de- after 13 sweeps contains only 1.9% misclassified pixels.
482 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486

Fig. 5. Texture mosaic C2.

Fig. 6. Texture mosaic C3.

Fig. 7. Texture mosaic C4.

Mosaics C4 (Fig. 7a) and C5 (Fig. 8a) are composed Our method achieves a smaller misclassification error of
of textures taken from outdoor scenes [20]. In their 1.4% (Fig. 7c). The difference between MIR (4.1) and

study, Jain and Karu tackled the problem of texture MIR (1.3) is considerable, which reflects the reliability

segmentation with a neural network generalization of the of the analysis.
traditional multichannel filtering method, using various For mosaic C5, which is 384;384 pixels in size, Jain
filter banks for feature extraction. For mosaic C4, which and Karu reported a labeling error of 6% with Laws’
is 256;256 pixels in size, they obtained a segmentation filters in supervised mode. Our unsupervised method
error of 3.3% with learned masks in unsupervised mode. gives a clearly better segmentation result of 2.1%. Note
Their method was not strictly unsupervised, though, be- that the pixelwise classification clearly improves the re-
cause the number of clusters was manually set to five. sult of the agglomerative merging phase (7.8%). The
 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486 483

Fig. 8. Texture mosaic C5.

Fig. 9. Segmentation of a mosaic which contains two different orientations of a directional texture.

difference between MIR (2.8) and MIR (1.2) is still can observe from the image, the textures of natural scenes
 
noticeable, but by far the smallest in the three cases, are generally more nonuniform than the homogeneous
reflecting the inherent difficulty of this problem. textures of the test mosaics. Also, in natural scenes adjac-
LBP/C transform is by definition rotation-variant. ent textured regions are not necessarily separated by
Fig. 9 demonstrates how the segmentation algorithm well-defined boundaries, but the spatial pattern smoothly
works in the case of an edge between two different ori- changes from one texture to another. Further, we have to
entations of a directional texture. The original texture observe the infinite scale of texture differences present in
(D21 from the Brodatz album) was rotated 30° in both natural scenes; choosing the right scale is a very subjec-
clockwise and counterclockwise direction using cubic in- tive matter. For these reasons there is often no ‘correct’
terpolation, and the rotated textures were merged into the segmentation for a natural scene, as is the case with
200;200 mosaic shown in Fig. 9a. This size guarantees texture mosaics.
that the edge between the two orientations is not acciden- The parameters X and ½ primarily control the scale of
tally aligned with the initial blocks which could bias the texture differences that will be detected. With values
result. The same set of parameter values was used as with X"1.1 and ½"1.5 the rough segmentation results after
texture mosaics C1 to C5. The final segmentation result the agglomerative merging phase are presented in Figs.
in Fig. 9b contains 66 mislabelled pixels along the edge. 10b and 11b, and the final segmentation results are
We also applied the texture segmentation method to shown in Figs. 10c and 11c, respectively. If we decreased
natural scenes. The scenes were originally in RGB format ½ further, the segmentation result would contain an
[25], but we converted them to gray-level intensity im- increasing number of regions. The invariance of the
ages. As an example, scene C1 (Fig. 10a) is a 384;384 LBP/C transform to average gray-level shows in the
image of rocks in the sea and scene C2 (Fig. 11a) is bottom part of the image, where the sea is interpreted as
a 192;192 image of a beach, water and foliage. As we a single region despite the shadows.
484 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486

Fig. 10. Natural scene C1.

Fig. 11. Natural scene C2.

The results obtained for these natural scenes are very ther, we could consider a particular feature at multiple
satisfactory, considering that important color or gray scales, by straightforwardly computing the desired fea-
scale information is not utilized in the segmentation. ture for suitably symmetrical discrete neighborhoods of
any size, such as disks or boxes of odd or even size.
A simple way to define a ‘‘multiresolution’’ LBP would
5. Discussion be to choose the eight neighbors of the center pixel from
the corresponding positions in different neighborhoods
In the presented method texture is described by joint (3;3, 5;5, 7;7, etc.).
occurrences of LBP and C. Obvious generalizations are The remaining question is how to combine the mul-
to use other texture features or feature domains (e.g. tiple feature channels obtained with several features
color) and scale. Although LBP/C is a very powerful and/or scales. We can hardly expect to reliably estimate
texture transform, we expect to achieve better results by joint distributions for a large number of features. Also,
combining a larger number of features in the analysis. multidimensional histograms with large numbers of bins
Other powerful texture measures, like distributions based are very computationally intensive and consume very
on gray-level difference histograms or co-occurrence ma- much memory. An alternative is to use an approximation
trices, can be easily incorporated into our algorithm. In with marginal distributions and to employ each indepen-
Pietikäinen, Nieminen, Marszalec and Ojala [33] we dent feature separately, as a 1-D histogram, to compute
demonstrated that a method based on comparison of a similarity score such as G for each feature, and then
feature distributions can be used for high-accuracy color integrate individual scores into an aggregate similarity
measurements. This suggests that distributions of color score. This approach has given very promising results in
features could be easily used to find small color differ- our texture classification experiments [29] and more
ences between neighboring regions in segmentation. recently in color classification as well [33]. Combining it
Color features should make our method efficient for with a carefully chosen set of non-redundant com-
segmenting images containing color textures, like the plementary features we expect to improve the perfor-
original color images used in the experiments [25]. Fur- mance of our segmentation method considerably. In
 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486 485

a similar way, joint pairs of features, like LBP/C, can be for providing images used in this study; Richard C.
combined with other single features or feature pairs. It Dubes, Anil K. Jain, John Lees and Kalle Karu from the
would also be possible to use single features or joint Michigan State University and Glenn Healey and David
features one by one, by e.g. first comparing the uniform- Slater from the University of California at Irvine.
ity of regions with respect to texture and then with
respect to color.
The histogram comparison approach based on the References
G test could be replaced with some other related method,
like histogram intersection [34] or a statistical chi-square [1] M. Tuceryan, A.K. Jain, Texture analysis, in Handbook
test. According to our experience the choice of proper of Pattern Recognition and Computer Vision, C.H. Chen,
texture measures is usually a much more important fac- L. F. Pau, P.S.P. Wang, eds., Ch. 2.1, pp. 235—276. World
tor in texture discrimination than the particular method Scientific, Singapore, 1993, pp. 235—276.
used for histogram comparison [29]. However, it would [2] R.M. Haralick, L.G. Shapiro, Image segmentation tech-
niques, Computer Vision Graphics, Image Processing 29
be interesting to study the performances of different types
(1985) 100—132.
of approaches in the case of very small image windows.
[3] T.R. Reed, J.M.H. Du Buf, A review of recent texture
In recent texture classification studies [35], we have segmentation, feature extraction techniques, CVGIP
compared the performance of LBP and other operators Image Understanding 57 (1993) 359—372.
to that of GMRF and Gabor energy features. The image [4] N.R. Pal, S.K. Pal, A review on image segmentation tech-
data included both Brodatz textures and the many differ- niques, Pattern Recognition 26 (1993) 1277—1294.
ent texture images available at MeasTex [36] which is an [5] P.C. Chen, T. Pavlidis, Segmentation by texture using
independent texture classification algorithm evaluation a co-occurrence matrix and a split-and-merge algorithm,
site accessible in WWW. LBP did well in these classifica- Comput. Graphics Image Processing 10 (1979) 172—182.
tion experiments, hence it should be suitable for texture [6] M. Pietikäinen, A. Rosenfeld, Image segmentation by tex-
ture using pyramid node linking, IEEE Trans. Systems
segmentation as well.
Man Cybernet. 11 (1981) 822—825.
[7] L.S. Davis, A. Mitiche, Mites: A model driven, iterative
texture segmentation algorithm, Comput. Graphics Image
6. Conclusion Processing 19 (1982) 95—110.
[8] M. Spann, R. Wilson, A quad-tree approach to image
We proposed a solution to unsupervised texture seg- segmentation which combines statistical and spatial in-
mentation, in which a method based on comparison of formation, Pattern Recognition 18 (1985) 257—269.
feature distributions is used to find homogeneously tex- [9] M. Unser, M. Eden, Multiresolution feature extraction
tured image regions and to localize boundaries between and selection for texture segmentation, IEEE Trans. Pat-
regions. Texture information is measured with a method tern Anal. Machine Intelligence 11 (1989) 717—728.
based on local binary patterns and contrast (LBP/C) that [10] J.Y. Hsiao, A.A. Sawchuk, Unsupervised texture image
we have recently developed. A region-based algorithm is segmentation using feature smoothing and probabilistic
relaxation techniques, Comput. Vision Graphics Image
developed for coarse image segmentation and a pixelwise
Processing 48 (1989) 1—21.
classification scheme for improving the localization of
[11] A. Khotanzad, A. Bouarfa, A parallel non-parametric clus-
region boundaries. tering algorithm with application to image segmentation,
The method performed very well in experiments. It is Proc. 22nd Asilomar Conf. on Signals, Systems, and Com-
not sensitive to the selection of parameter values, does puters, pp. 305—309, Pacific Grove, CA 1988.
not require any prior knowledge about the number of [12] J. Mao, A.K. Jain, Texture classification and segmentation
textures or regions in the image, and seems to provide using multiresolution simultaneous autoregressive models,
significantly better results than existing unsupervised tex- Pattern Recognition 25 (1992) 173—188.
ture segmentation approaches. The method can be easily [13] J.F. Silverman, D.B. Cooper, Bayesian clustering for un-
generalized, e.g. to utilize other texture features, multi- supervised estimation of surface and texture models, IEEE
scale information, color features, and combinations of Trans. Pattern Anal. Machine Intelligence 10 (1988)
multiple features. 482—495.
[14] C. Bouman, B. Liu, Multiple resolution segmentation of
textured images, IEEE Trans. Pattern Anal. Machine In-
telligence 13 (1991) 99—113.
Acknowledgements
[15] B.S. Manjunath, R. Chellappa, Unsupervised texture seg-
mentation using Markov random field models, IEEE
The financial support provided by the Academy of Trans. on Pattern Anal. Machine Intelligence 13 (1991)
Finland, the Technology Development Center of 478—482.
Finland, and the Graduate School in Electronics, Tele- [16] F.S. Cohen, Z. Fan, Maximum likelihood unsupervised
communications and Automation is gratefully acknow- textured image segmentation, CVGIP: Graphical Models
ledged. The authors also wish to thank following persons and Image Processing 54 (1992) 239—251.
486 T. Ojala, M. Pietikäinen / Pattern Recognition 32 (1999) 477—486

[17] A. Bovik, M. Clark, W.S. Geisler, Multichannel texture [27] T. Ojala, M. Pietikäinen, D. Harwood, A comparative
analysis using localized spatial filters, IEEE Trans. Pattern study of texture measures with classification based on
Anal. and Machine Intelligence 12 (1990) 55—73. feature distributions, Pattern Recognition 29 (1996) 51—59.
[18] A.K. Jain, F. Farrokhnia, Unsupervised texture segmenta- [28] T. Ojala, M. Pietikäinen, J. Nisula, Determining composi-
tion using Gabor filters, Pattern Recognition 24 (1991) tion of grain mixtures by texture classification based on
1167—1186. feature distributions, Int. J. Pattern Recognition Artif.
[19] H. Greenspan, R. Goodman, R. Chellappa, C.H. Ander- Intelligence 10 (1996) 73—82.
son, Learning texture discrimination rules in a multi- [29] T. Ojala, Multichannel approach to texture description
resolution system, IEEE Trans. Pattern Anal. Machine with feature distributions, Technical Report CAR-TR-846,
Intelligence 16 (1994) 894—901. Center for Automation Research, University of Maryland,
[20] A.K. Jain, K. Karu, Learning texture discrimination 1996.
masks, IEEE Trans. Pattern Anal. Machine Intelligence [30] R.R. Sokal, F.J. Rohlf, Introduction to Biostatistics, 2nd
18 (1996) 195—205. edn. W.H. Freeman and Co, New York, 1987.
[21] M. Unser, Texture classification and segmentation using [31] P. Brodatz, Textures: A Photographic Album for Artists
wavelet frames, IEEE Trans. Image Processing 4 (1995) and Designers. Dover, New York, 1966.
1549—1560. [32] P.P. Ohanian, R.C. Dubes, Performance evaluation for
[22] A. Laine, J. Fan, Frame representations for texture seg- four classes of textural features, Pattern Recognition 25
mentation, IEEE Trans. Image Processing 5 (1996) (1992) 819—833.
771—780. [33] M. Pietikäinen, S. Nieminen, E. Marszalec, T. Ojala,
[23] B.B. Chaudhuri, N. Sarkar, Texture segmentation using Accurate color discrimination with classification based
fractal dimension, IEEE Trans. Pattern Anal. Machine on feature distributions, Proc. 13th Int. Conf. on
Intelligence 17 (1995) 72—77. Pattern Recognition, Vol. 3, pp. 833—838, Vienna, Austria,
[24] J.-L. Chen, A. Kundu, Unsupervised texture segmentation 1996.
using multichannel decomposition and hidden Markov [34] M. Swain, D. Ballard, Color indexing, Int. J. Computer
models, IEEE Trans. Image Processing 4 (1995) 603—619. Vision 7 (1991) 11—32.
[25] D.K. Panjwani, G. Healey, Markov random field models [35] T. Ojala, Nonparametric texture analysis using spatial
for unsupervised segmentation of textured color images, operators, with applications in visual inspection, Ph.D.
IEEE Trans. Pattern Anal. Machine Intelligence 17 (1995) thesis, Acta Univ. Oul. C 105, University of Oulu, Finland,
939—954. 1997.
[26] D. Harwood, T. Ojala, M. Pietikäinen, S. Kelman, [36] G. Smith, I. Burns, MeasTex Image Texture Database
L.S. Davis, Texture classification by center-symmetric and Test Suite, CSSIP, University of Queensland,
auto-correlation, using Kullback discrimination of distri- Australia. http://www.cssip.elec.uq.edu.au/)guy/meastex/
butions, Pattern Recognition Lett. 16 (1995) 1—10. home.html.

About the Author—TIMO OJALA received the M.S. degree in Electrical Engineering with honors from the University of Oulu, Finland,
in 1992, and the Dr. Tech. degree from the same university in 1997. He is a member of Machine Vision and Media Processing Group at
the University of Oulu. From 1996 to 1997 he was visiting the University of Maryland Institute for Advanced Computer Studies
(UMIACS). His research interests includes pattern recognition, texture analysis and object-oriented software design.

About the Author—MATTI PIETIKA®INEN received his Doctor of Technology degree in Electrical Engineering from the University of
Oulu, Finland, in 1982. Currently he is Professor of Information Technology, Scientific Director of Infotech Oulu—a center for
information technology research, and Head of Machine Vision and Media Processing Group, at the University of Oulu. From 1980 to
1981 and from 1984 to 1985 he was visiting the Computer Vision Laboratory at the University of Maryland, USA. His research interests
include machine vision, document analysis, and their applications. He has authored about 100 papers in journals, books and
conferences. He is the editor (with L.F. Pau) of the book ‘‘Machine Vision for Advanced Production’’, published by World Scientific in
1996. Prod. Pietikäinen is Fellow of International Association for Pattern Recognition (IAPR) and Senior Member of IEEE, and serves
as Member of the Governing Board of IAPR and Chairman of IAPR Education Committee. He also serves on program committees of
several international conferences.