International Journal of Computer Vision 43(1), 727, 2001

c 2001 Kluwer Academic Publishers. Manufactured in The Netherlands.



Contour and Texture Analysis for Image Segmentation

JITENDRA MALIK, SERGE BELONGIE, THOMAS LEUNG AND JIANBO SHI
Computer Science Division, University of California at Berkeley, Berkeley, CA 94720-1776, USA
Received December 28, 1999; Revised February 23, 2001; Accepted February 23, 2001

Abstract. This paper provides an algorithm for partitioning grayscale images into disjoint regions of coherent
brightness and texture. Natural images contain both textured and untextured regions, so the cues of contour and
texture differences are exploited simultaneously. Contours are treated in the intervening contour framework, while
texture is analyzed using textons. Each of these cues has a domain of applicability, so to facilitate cue combination we
introduce a gating operator based on the texturedness of the neighborhood at a pixel. Having obtained a local measure
of how likely two nearby pixels are to belong to the same region, we use the spectral graph theoretic framework of
normalized cuts to find partitions of the image into regions of coherent texture and brightness. Experimental results
on a wide range of images are shown.
Keywords: segmentation, texture, grouping, cue integration, texton, normalized cut
1.

Introduction

To humans, an image is not just a random collection

of pixels; it is a meaningful arrangement of regions
and objects. Figure 1 shows a variety of images. Despite the large variations of these images, humans have
no problem interpreting them. We can agree about the
different regions in the images and recognize the different objects. Human visual grouping was studied extensively by the Gestalt psychologists in the early part of
the 20th century (Wertheimer, 1938). They identified
several factors that lead to human perceptual grouping: similarity, proximity, continuity, symmetry, parallelism, closure and familiarity. In computer vision,
these factors have been used as guidelines for many
grouping algorithms.
The most studied version of grouping in computer vision is image segmentation. Image segmentation techniques can be classified into two broad families
(1) region-based, and (2) contour-based approaches.
Region-based approaches try to find partitions of the
image pixels into sets corresponding to coherent im Present
Present

address: Compaq Cambridge Research Laboratory.

address: Robotics Institute, Carnegie Mellon University.

age properties such as brightness, color and texture.

Contour-based approaches usually start with a first
stage of edge detection, followed by a linking process
that seeks to exploit curvilinear continuity.
These two approaches need not be that different from
each other. Boundaries of regions can be defined to be
contours. If one enforces closure in a contour-based
framework (Elder and Zucker, 1996; Jacobs, 1996)
then one can get regions from a contour-based approach. The difference is more one of emphasis and
what grouping factor is coded more naturally in a given
framework.
A second dimension on which approaches can
be compared is local vs. global. Early techniques,
in both contour and region frameworks, made local
decisionsin the contour framework this might be
declaring an edge at a pixel with high gradient, in the
region framework this might be making a merge/split
decision based on a local, greedy strategy.
Region-based techniques lend themselves more
readily to defining a global objective function (for
example, Markov random fields (Geman and Geman, 1984) or variational formulations (Mumford and
Shah, 1989)). The advantage of having a global objective function is that decisions are made only when

Malik et al.

Figure 1. Some challenging images for a segmentation algorithm. Our goal is to develop a single grouping procedure which can deal with all
these types of images.

information from the whole image is taken into account

at the same time.
In contour-based approaches, often the first step of
edge detection is done locally. Subsequently efforts are
made to improve results by a global linking process that
seeks to exploit curvilinear continuity. Examples include dynamic programming (Montanari, 1971), relaxation approaches (Parent and Zucker, 1989), saliency
networks (Shaashua and Ullman, 1988), stochastic
completion (Williams and Jacobs, 1995). A criticism
of this approach is that the edge/no edge decision is
made prematurely. To detect extended contours of very
low contrast, a very low threshold has to be set for the
edge detector. This will cause random edge segments
to be found everywhere in the image, making the task
of the curvilinear linking process unnecessarily harder
than if the raw contrast information was used.
A third dimension on which various segmentation
schemes can be compared is the class of images for
which they are applicable. As suggested by Fig. 1, we
have to deal with images which have both textured and
untextured regions. Here boundaries must be found using both contour and texture analysis. However what
we find in the literature are approaches which concentrate on one or the other.
Contour analysis (e.g. edge detection) may be adequate for untextured images, but in a textured region

it results in a meaningless tangled web of contours.

Think for instance of what an edge detector would return on the snow and rock region in Fig. 1(a). The
traditional solution for this problem in edge detection is to use a high threshold so as to minimize the
number of edges found in the texture area. This is obviously a non-solutionsuch an approach means that
low-contrast extended contours will be missed as well.
This problem is illustrated in Fig. 2. There is no recognition of the fact that extended contours, even weak in
contrast, are perceptually significant.
While the perils of using edge detection in textured
regions have been noted before (see e.g. Binford, 1981),
a complementary problem of contours constituting a
problem for texture analysis does not seem to have been
recognized before. Typical approaches are based on
measuring texture descriptors over local windows, and
then computing differences between window descriptors centered at different locations. Boundaries can then
give rise to thin strip-like regions, as in Fig. 3. For specificity, assume that the texture descriptor is a histogram
of linear filter outputs computed over a window. Any
histogram window near the boundary of the two regions
will contain large filter responses from filters oriented
along the direction of the edge. However, on both sides
of the boundary, the histogram will indicate a featureless region. A segmentation algorithm based on, say, 2

Contour and Texture Analysis

Figure 2. Demonstration of texture as a problem for the contour process. Each image shows the edges found with a Canny edge detector for the
penguin image using different scales and thresholds: (a) fine scale, low threshold, (b) fine scale, high threshold, (c) coarse scale, low threshold,
(d) coarse scale, high threshold. A parameter setting that preserves the correct edges while suppressing spurious detections in the textured area
is not possible.

Figure 3. Demonstration of the contour-as-a-texture problem using a real image. (a) Original image of a bald eagle. (b) The groups found
by an EM-based algorithm (Belongie et al., 1998).

distances between histograms, will inevitably partition

the boundary as a group of its own. As is evident, the
problem is not confined to the use of a histogram of filter outputs as texture descriptor. Figure 3(b) shows the
actual groups found by an EM-based algorithm using
an alternative color/texture descriptor (Belongie et al.,
1998).
1.1.

Desiderata of a Theory of Image Segmentation

At this stage, we are ready to summarize our desired

attributes for a theory of image segmentation.
1. It should deal with general images. Regions with
or without texture should be processed in the same

framework, so that the cues of contour and texture

differences can be simultaneously exploited.
2. In terms of contour, the approach should be able
to deal with boundaries defined by brightness step
edges as well as lines (as in a cartoon sketch).
3. Image regions could contain texture which could be
regular such as the polka dots in Fig. 1(c), stochastic
as in the snow and rock region in (a) or anywhere
in between such as the tiger stripes in (b). A key
question here is that one needs an automatic procedure for scale selection. Whatever ones choice
of texture descriptor, it has to be computed over a
local window whose size and shape need to be determined adaptively. What makes scale selection a
challenge is that the technique must deal with the

10

Malik et al.

wide range of texturesregular, stochastic, or intermediate casesin a seamless way.

1.2.

Introducing Textons

Julesz introduced the term texton, analogous to a

phoneme in speech recognition, nearly 20 years ago
(Julesz, 1981) as the putative units of preattentive human texture perception. He described them qualitatively for simple binary line segment stimulioriented
segments, crossings and terminatorsbut did not provide an operational definition for gray-level images.
Subsequently, texton theory fell into disfavor as a
model of human texture discrimination as accounts
based on spatial filtering with orientation and scaleselective mechanisms that could be applied to arbitrary
gray-level images became popular.
There is a fundamental, well recognized, problem
with linear filters. Generically, they respond to any
stimulus. Just because you have a response to an oriented odd-symmetric filter doesnt mean there is an
edge at that location. It could be that there is a higher
contrast bar at some other location in a different orientation which has caused this response. Tokens such as
edges or bars or corners can not be associated with the
output of a single filter. Rather it is the signature of the
outputs over scales, orientations and order of the filter
that is more revealing.
Here we introduce a further step by focussing on the
outputs of these filters considered as points in a high
dimensional space (on the order of 40 filters are used).
We perform vector quantization, or clustering, in this
high-dimensional space to find prototypes. Call these
prototypes textonswe will find empirically that these
tend to correspond to oriented bars, terminators and so
on. One can construct a universal texton vocabulary
by processing a large number of natural images, or
we could find them adaptively in windows of images.
In each case the K -means technique can be used. By
mapping each pixel to the texton nearest to its vector of
filter responses, the image can be analyzed into texton
channels, each of which is a point set.
It is our opinion that the analysis of an image into textons will prove useful for a wide variety of visual processing tasks. For instance, in Leung and Malik (1999)
we use the related notion of 3D textons for recognition
of textured materials. In the present paper, our objective is to develop an algorithm for the segmentation
of an image into regions of coherent brightness and
texturewe will find that the texton representation will

enable us to address the key problems in a very natural

fashion.
1.3.

Summary of Our Approach

We pursue image segmentation in the framework of

Normalized Cuts introduced by Shi and Malik (1997,
2000). The image is considered to be a weighted graph
where the nodes i and j are pixels and edge weights,
Wij , denote a local measure of similarity between the
two pixels. Grouping is performed by finding eigenvectors of the Normalized Laplacian of this graph (3). The
fundamental issue then is that of specifying the edge
weights Wij ; we rely on normalized cuts to go from
these local measures to a globally optimal partition of
the image.
The algorithm analyzes the image using the two cues
of contour and texture. The local similarity measure
between pixels i and j due to the contour cue, WijIC ,
is computed in the intervening contour framework of
Leung and Malik (1998) using peaks in contour orientation energy (2 and 4.1). Texture is analysed using textons (2.1). Appropriate local scale is estimated
from the texton labels. A histogram of texton densities is used as the texture descriptor. Similarity, WijTX ,
is measured using the 2 test on the histograms (4.2).
The edge weights Wij combining both contour and texture information are specified by gating each of the two
cues with a texturedness measure (4.3).
In (5), we present the practical details of going from
the eigenvectors of the normalized Laplacian matrix of
the graph to a partition of the image. Results from the
algorithm are presented in (6). Some of the results
presented here were published in Malik et al. (1999).

2.

Filters, Composite Edgels, and Textons

Since the 1980s, many approaches have been proposed

in the computer vision literature that start by convolving the image with a bank of linear spatial filters f i
tuned to various orientation and spatial frequencies
(Knutsson and Granlund, 1983; Koenderink and van
Doorn, 1987; Fogel and Sagi, 1989; Malik and Perona,
1990). (See Fig. 4 for an example of such a filter
set.)
These approaches were inspired by models of processing in the early stages of the primate visual system
(e.g. DeValois and DeValois, 1988). The filter kernels
f i are models of receptive fields of simple cells in visual

Contour and Texture Analysis

11

Figure 4. Left: Filter set f i consisting of 2 phases (even and odd), 3 scales (spaced by half-octaves), and 6 orientations (equally spaced from
0 to ). The basic filter is a difference-of-Gaussian quadrature pair with 3 : 1 elongation. Right: 4 scales of center-surround filters. Each filter is
L 1 -normalized for scale invariance.

cortex. To a first approximation, we can classify them

into three categories:
1. Cells with radially symmetric receptive fields. The
usual choice of f i is a Difference of Gaussians
(DOG) with the two Gaussians having different values of . Alternatively, these receptive fields can
also be modeled as the Laplacian of Gaussian.
2. Oriented odd-symmetric cells whose receptive
fields can be modeled as rotated copies of a horizontal oddsymmetric receptive field. A suitable
point spread function for such a receptive field is
f (x, y) = G 1 (y)G 2 (x) where G (x) represents
a Gaussian with standard deviation . The ratio
2 : 1 is a measure of the elongation of the filter.
3. Oriented even-symmetric cells whose receptive
fields can be modeled as rotated copies of a horizontal evensymmetric receptive field. A suitable point
spread function for such a receptive field is
f (x, y) = G 1 (y)G 2 (x)
The use of Gaussian derivatives (or equivalently, differences of offset Gaussians) for modeling receptive
fields of simple cells is due to Young (1985). One could
equivalently use Gabor functions. Our preference for
Gaussian derivatives is based on their computational
simplicity and their natural interpretation as blurred
derivatives (Koenderink and van Doorn, 1987, 1988).
The oriented filterbank used in this work, depicted
in Fig. 4, is based on rotated copies of a Gaussian
derivative and its Hilbert transform. More precisely,
let f 1 (x, y) = G 1 (y)G 2 (x) and f 2 (x, y) equal the

Hilbert transform of f 1 (x, y) along the y axis:


 2 
 2
1
x
d2
y
f 1 (x, y) =
exp 2 exp 2 2
dy 2 C


f 2 (x, y) = Hilbert( f 1 (x, y))
where is the scale,  is the aspect ratio of the filter, and C is a normalization constant. (The use of the
Hilbert transform instead of a first derivative makes f 1
and f 2 an exact quadrature pair.) The radially symmetric portion of the filterbank consists of Difference-ofGaussian kernels. Each filter is zero-mean and L 1 normalized for scale invariance (Malik and Perona, 1990).
Now suppose that the image is convolved with such
a bank of linear filters. We will refer to the collection of
response images I f i as the hypercolumn transform
of the image.
Why is this useful from a computational point of
view? The vector of filter outputs I f i (x0 , y0 ) characterizes the image patch centered at x0 , y0 by a set
of values at a point. This is similar to characterizing
an analytic function by its derivatives at a pointone
can use a Taylor series approximation to find the values of the function at neighboring points. As pointed
out by Koenderink and van Doorn (1987), this is more
than an analogy, because of the commutativity of the
operations of differentiation and convolution, the receptive fields described above are in fact computing
blurred derivatives. We recommend Koenderink and
van Doorn (1987, 1988), Jones and Malik (1992), and
Malik and Perona (1992) for a discussion of other advantages of such a representation.
The hypercolumn transform provides a convenient
front end for contour and texture analysis:

12

Malik et al.

Contour. In computational vision, it is customary

to model brightness edges as step edges and to detect them by marking locations corresponding to
the maxima of the outputs of odd-symmetric filters
(e.g. Canny, 1986) at appropriate scales. However,
it should be noted that step edges are an inadequate
model for the discontinuities in the image that result from the projection of depth or orientation discontinuities in physical scene. Mutual illumination
and specularities are quite common and their effects are particularly significant in the neighborhood of convex or concave object edges. In addition, there will typically be a shading gradient on
the image regions bordering the edge. As a consequence of these effects, real image edges are not
step functions but more typically a combination of
steps, peak and roof profiles. As was pointed out
in Perona and Malik (1990), the oriented energy
approach (Knutsson and Granlund, 1983; Morrone
and Owens, 1987; Morrone and Burr, 1988) can be
used to detect and localize correctly these composite edges.
The oriented energy, also known as the quadrature energy, at angle 0 is defined as:
OE0 = (I f 1 )2 + (I f 2 )2
OE0 has maximum response for horizontal contours. Rotated copies of the two filter kernels are
able to pick up composite edge contrast at various
orientations.
Given OE , we can proceed to localize the composite edge elements (edgels) using oriented nonmaximal suppression. This is done for each scale
in the following way. At a generic pixel q, let
= arg max OE denote the dominant orientation
and OE the corresponding energy. Now look at
the two neighboring values of OE on either side
of q along the line through q perpendicular to the
dominant orientation. The value OE is kept at the
location of q only if it is greater than or equal to
each of the neighboring values. Otherwise it is replaced with a value of zero.
Noting that OE ranges between 0 and infinity,
we convert it to a probability-like number between
0 and 1 as follows:
pcon = 1 exp(OE/IC )

(1)

IC is related to oriented energy response purely

due to image noise. We use = 0.02 in this paper.

The idea is that for any contour with OE IC ,

pcon 1.
Texture. As the hypercolumn transform provides a
good local descriptor of image patches, the boundary between differently textured regions may be
found by detecting curves across which there is a
significant gradient in one or more of the components of the hypercolumn transform. For an elaboration of this approach, see Malik and Perona
(1990).
Malik and Perona relied on averaging with large
kernels to smooth away spatial variation for filter
responses within regions of texture. This process
loses a lot of information about the distribution of
filter responses; a much better method is to represent the neighborhood around a pixel by a histogram of filter outputs (Heeger and Bergen, 1995;
Puzicha et al., 1997). While this has been shown to
be a powerful technique, it leaves open two important questions. Firstly, there is the matter of what
size window to use for pooling the histogramthe
integration scale. Secondly, these approaches only
make use of marginal binning, thereby missing out
on the informative characteristics that joint assemblies of filter outputs exhibit at points of interest.
We address each of these questions in the following
section.
2.1.

Textons

Though the representation of textures using filter responses is extremely versatile, one might say that it is
overly redundant (each pixel value is represented by
Nfil real-valued filter responses, where Nfil is 40 for our
particular filter set). Moreover, it should be noted that
we are characterizing textures, entities with some spatially repeating properties by definition. Therefore, we
do not expect the filter responses to be totally different at each pixel over the texture. Thus, there should
be several distinct filter response vectors and all others
are noisy variations of them.
This observation leads to our proposal of clustering the filter responses into a small set of prototype
response vectors. We call these prototypes textons. Algorithmically, each texture is analyzed using the filter
bank shown in Fig. 4. Each pixel is now transformed
to a Nfil dimensional vector of filter responses. These
vectors are clustered using K -means. The criterion for
this algorithm is to find K centers such that after assigning each data vector to the nearest center, the sum

Contour and Texture Analysis

13

Figure 5. (a) Polka-dot image. (b) Textons found via K -means with K = 25, sorted in decreasing order by norm. (c) Mapping of pixels to the
texton channels. The dominant structures captured by the textons are translated versions of the dark spots. We also see textons corresponding
to faint oriented edge and bar elements. Notice that some channels contain activity inside a textured region or along an oriented contour and
nowhere else.

of the squared distance from the centers is minimized.

K -means is a greedy algorithm that finds a local minimum of this criterion.1
It is useful to visualize the resulting cluster centers
in terms of the original filter kernels. To do this, recall
that each cluster center represents a set of projections of
each filter onto a particular image patch. We can solve
for the image patch corresponding to each cluster center
in a least squares sense by premultiplying the vectors
representing the cluster centers by the pseudoinverse of
the filterbank (Jones and Malik, 1992). The matrix rep-

resenting the filterbank is formed by concatenating the

filter kernels into columns and placing these columns
side by side. The set of synthesized image patches for
two test images are shown in Figs. 5(b) and 6(b). These
are our textons. The textons represent assemblies of
filter outputs that are characteristic of the local image
structure present in the image.
Looking at the polka-dot example, we find that many
of the textons correspond to translated versions of dark
spots.2 Also included are a number of oriented edge
elements of low contrast and two textons representing

14

Malik et al.

Figure 6. (a) Penguin image. (b) Textons found via K -means with K = 25, sorted in decreasing order by norm. (c) Mapping of pixels to the
texton channels. Among the textons we see edge elements of varying orientation and contrast along with elements of the stochastic texture in
the rocks.

nearly uniform brightness. The pixel-to-texton mapping is shown in Fig. 5(c). Each subimage shows the
pixels in the image that are mapped to the corresponding texton in Fig. 5(b). We refer to this collection of
discrete point sets as the texton channels. Since each
pixel is mapped to exactly one texton, the texton channels constitute a partition of the image.
Textons and texton channels are also shown for the
penguin image in Fig. 6. Notice in the two examples
how much the texton set can change from one image

to the next. The spatial characteristics of both the deterministic polka dot texture and the stochastic rocks
texture are captured across several texton channels. In
general, the texture boundaries emerge as point density
changes across the different texton channels. In some
cases, a texton channel contains activity inside a particular textured region and nowhere else. By comparison, vectors of filter outputs generically respond with
some value at every pixela considerably less clean
alternative.

Contour and Texture Analysis

We have not been particularly sophisticated in the

choice of K , the number of different textons for a given
image. How to choose an optimal value of K in K means has been the subject of much research in the
model selection and clustering literature; we used a
fixed choice K = 36 to obtain the segmentation results
in this paper. Clearly, if the images vary considerably in
complexity and number of objects in them, an adaptive
choice may give better results.
The mapping from pixel to texton channel provides
us with a number of discrete point sets where before
we had continuous-valued filter vectors. Such a representation is well suited to the application of techniques
from computational geometry and point process statistics. With these tools, one can approach questions such
as, what is the neighborhood of a texture element?
and how similar are two pixels inside a textured
region?
Several previous researchers have employed clustering using K -means or vector quantization as a stage in
their approach to texture classificationtwo representative examples are McLean (1993) and Raghu et al.
(1997). What is novel about our approach is the identification of clusters of vectors of filter outputs with the
Julesz notion of textons. Then first order statistics of
textons are used for texture characterization, and the
spatial structure within texton channels enables scale
estimation. Vector quantization becomes much more
than just a data compression or coding step. The next
subsection should make this point clear.
2.1.1. Local Scale and Neighborhood Selection. The
texton channel representation provides us a natural way
to define texture scale. If the texture is composed of discrete elements(texels), we might want to define a notion of texel neighbors and consider the mean distance

15

between them to be a measure of scale. Of course, many

textures are stochastic and detecting texels reliably is
hard even for regular textures.
With textons we have a soft way to define neighbors. For a given pixel in a texton channel, first consider it as a thickened point a disk centered at it.3
The idea is that while textons are being associated with
pixels, since they correspond to assemblies of filter outputs, it is better to think of them as corresponding to
a small image disk defined by the scale used in the
Gaussian derivative filters. Recall Koenderinks aphorism about a point in image analysis being a Gaussian
blob of small !
Now consider the Delaunay neighbors of all the pixels in the thickened point of a pixel i which lie closer
than some outer scale.4 The intuition is that these will
be pixels in spatially neighboring texels. Compute the
distances of all these pixels to i; the median of these
constitutes a robust local measure of inter-texel distance. We define the local scale (i) to be 1.5 times
this median distance.
In Fig. 7(a), the Delaunay triangulation of a zoomedin portion of one of the texton channels in the polka-dot
dress of Fig. 5(a) is shown atop a brightened version
of the image. Here the nodes represent points that are
similar in the image while the edges provide proximity
information.
The local scale (i) is based just on the texton channel for the texton at i. Since neighboring pixels should
have similar scale and could be drawn from other texton channels, we can improve the estimate of scale by
median filtering of the scale image.
2.1.2. Computing Windowed Texton Histograms.
Pairwise texture similarities will be computed by comparing windowed texton histograms. We define the

Figure 7. Illustration of scale selection. (a) Closeup of Delaunay triangulation of pixels in a particular texton channel for polka dot image. (b)
Neighbors of thickened point for pixel at center. The thickened point lies within inner circle. Neighbors are restricted to lie within outer circle.
(c) Selected scale based on median of neighbor edge lengths, shown by circle, with all pixels falling inside circle marked with dots.

16

Malik et al.

window W(i) for a generic pixel i as the axis-aligned

square of radius (i) centered on pixel i.
Each histogram has K bins, one for each texton channel. The value of the kth histogram bin for a pixel i is
found by counting how many pixels in texton channel k
fall inside the window W(i). Thus the histogram represents texton frequencies in a local neighborhood. We
can write this as
h i (k) =

I [T ( j) = k]

(2)

For more discussion of this criterion, please refer to Shi

and Malik (2000).
One key advantage of using the normalized cut is that
a good approximation to the optimal partition can be
computed very efficiently.5 Let W be the association
matrix, i.e. Wij is the weight between nodes i and j
in the 
graph. Let D be the diagonal matrix such that
Dii = j Wij , i.e. Dii is the sum of the weights of all
the connections to node i. Shi and Malik showed that
the optimal partition can be found by computing:

jW(i)

y = arg min Ncut

yT (D W)y
= arg min
y
yT Dy

where I [] is the indicator function and T ( j) returns

the texton assigned to pixel j.
3.

The Normalized Cut Framework

In the Normalized Cut framework (Shi and Malik,

1997, 2000), which is inspired by spectral graph theory
(Chung, 1997), Shi and Malik formulate visual grouping as a graph partitioning problem. The nodes of the
graph are the entities that we want to partition; for example, in image segmentation, they are the pixels. The
edges between two nodes correspond to the strength
with which these two nodes belong to one group; again,
in image segmentation, the edges of the graph correspond to how much two pixels agree in brightness,
color, etc. Intuitively, the criterion for partitioning the
graph will be to minimize the sum of weights of connections across the groups and maximize the sum of
weights of connections within the groups.
Let G = {V, E} be a weighted undirected graph,
where V are the nodes and E are the edges. Let A, B
be a partition of the graph: A B = V, A B = . In
graph theoretic language, the similarity between these
two groups is called the cut:
cut(A, B) =

Wij

iA, jB

where Wij is the weight on the edge between nodes

i and j. Shi and Malik proposed to use a normalized
similarity criterion to evaluate a partition. They call it
the normalized cut:
cut(A, B)
cut(B, A)
N cut(A, B) =
+
assoc(A, V) assoc(B, V)

where assoc(A, V) = iA,kV Wik is the total connection from nodes in A to all the nodes in the graph.

(3)

where y = {a, b} N is a binary indicator vector specifying the group identity for each pixel, i.e. yi = a if
pixel i belongs to group A and y j = b if pixel j belongs
to B. N is the number of pixels. Notice that the above
expression is a Rayleigh quotient. If we relax y to take
on real values (instead of two discrete values), we can
optimize Eq. (3) by solving a generalized eigenvalue
system. Efficient algorithms with polynomial running
time are well-known for solving such problems.
The process of transforming the vector y into a discrete bipartition and the generalization to more than
two groups is discussed in (5).
4.

Defining the Weights

The quality of a segmentation based on Normalized

Cuts or any other algorithm based on pairwise similarities fundamentally depends on the weightsthe
Wij sthat are provided as input. The weights should
be large for pixels that should belong together and small
otherwise. We now discuss our method for computing
the Wij s. Since we seek to combine evidence from
two cues, we will first discuss the computation of the
weights for each cue in isolation, and then describe
how the two weights can be combined in a meaningful
fashion.
4.1.

Images Without Texture

Consider for the moment the cracked earth image in

Fig. 1(e). Such an image contains no texture and may be
treated in a framework based solely on contour features.
The definition of the weights in this case, which we
denote WijIC , is adopted from the intervening contour
method introduced in Leung and Malik (1998).

Contour and Texture Analysis

17

Figure 8. Left: the original image. Middle: part of the image marked by the box. The intensity values at pixels p1 , p2 and p3 are similar.
However, there is a contour in the middle, which suggests that p1 and p2 belong to one group while p3 belongs to another. Just comparing
intensity values at these three locations will mistakenly suggest that they belong to the same group. Right: orientation energy. Somewhere along
l2 , the orientation energy is strong which correctly proposes that p1 and p3 belong to two different partitions, while orientation energy along l1
is weak throughout, which will support the hypothesis that p1 and p2 belong to the same group.

Figure 8 illustrates the intuition behind this idea. On

the left is an image. The middle figure shows a magnified part of the original image. On the right is the
orientation energy. There is an extended contour separating p3 from p1 and p2 . Thus, we expect p1 to be
much more strongly related to p2 than p3 . This intuition
carries over in our definition of dissimilarity between
two pixels: if the orientation energy along the line between two pixels is strong, the dissimilarity between
these pixels should be high (and Wij should be low).
Contour information in an image is computed
softly through orientation energy (OE) from elongated quadrature filter pairs. We introduce a slight modification here to allow for exact sub-pixel localization
of the contour by finding the local maxima in the orientation energy perpendicular to the contour orientation
(Perona and Malik, 1990). The orientation energy gives
the confidence of this contour. WijIC is then defined as
follows:
WijIC = 1 max pcon (x)
xMij

where Mij is the set of local maxima along the line joining pixels i and j. Recall from (2) that pcon (x), 0 <
pcon < 1, is nearly 1 whenever the orientated energy
maximum at x is sufficiently above the noise level. In
words, two pixels will have a weak link between them
if there is a strong local maximum of orientation energy
along the line joining the two pixels. On the contrary, if
there is little energy, for example in a constant brightness region, the link between the two pixels will be
strong. Contours measured at different scales can be
taken into account by computing the orientation energy maxima at various scales and setting pcon to be
the maximum over all the scales at each pixel.

4.2.

Images that are Texture Mosaics

Now consider the case of images wherein all of the

boundaries arise from neighboring patches of different
texture (e.g. Fig. 1(d)). We compute pairwise texture
similarities by comparing windowed texton histograms
computed using the technique described previously
(2.1.2). A number of methods are available for comparing histograms. We use the 2 test, defined as
2 (h i , h j ) =

K
[h i (k) h j (k)]2
1
2 k=1 h i (k) + h j (k)

where h i and h j are the two histograms. For an empirical comparison of the 2 test versus other texture
similarity measures, see Puzicha et al. (1997).
WijTX is then defined as follows:
WijTX = exp( 2 (h i , h j )/TX )

(4)

If histograms h i and h j are very different, 2 is large,

and the weight WijTX is small.
4.3.

General Images

Finally we consider the general case of images that

contain boundaries of both kinds. This presents us with
the problem of cue integration. The obvious approach
to cue integration is to define the weight between pixels
i and j as the product of the contribution from each
cue: Wij = WijIC WijTX . The idea is that if either of
the cues suggests that i and j should be separated,
the composite weight, Wij , should be small. We must
be careful, however, to avoid the problems listed in the

18

Malik et al.

Introduction (1) by suitably gating the cues. The spirit

of the gating method is to make each cue harmless
in locations where the other cue should be operating.
4.3.1. Estimating Texturedness. As illustrated in
Fig. 2, the fact that a pixel survives the non-maximum
suppression step does not necessarily mean that that
pixel lies on a region boundary. Consider a pixel inside
a patch of uniform texture: its oriented energy is large
but it does not lie on the boundary of a region. Conversely, consider a pixel lying between two uniform
patches of just slightly different brightness: it does lie
on a region boundary but its oriented energy is small.
In order to estimate the probability that a pixel lies
on a boundary, it is necessary to take more surrounding information into account. Clearly the true value
of this probability is only determined after the final
correct segmentation, which is what we seek to find.
At this stage our goal is to formulate a local estimate
of the texturedness of the region surrounding a pixel.
Since this is a local estimate, it will be noisy but its
objective will be to bootstrap the global segmentation
procedure.
Our method of computing this value is based on a
simple comparison of texton distributions on either side
of a pixel relative to its dominant orientation. Consider
a generic pixel q at an oriented energy maximum. Let
the dominant orientation be . Consider a circle of radius (q) (the selected scale) centered on q. We first
divide this circle in two along the diameter with orientation . Note that the contour passing through q
is tangent to the diameter, which is its best straight
line approximation. The pixels in the disk can be partitioned into three sets D0 , D , D+ which are the pixels
in the strip along the diameter, the pixels to the left
of D0 , and the pixels to the right of D0 , respectively.
To compute our measure of texturedness, we consider
two half window comparisons with D0 assigned to
each side. Assume without loss of generality that D0
is first assigned to the left half. Denote the K -bin
histograms of D0 D by h L and D+ by h R respectively. Now consider the 2 statistic between the two
histograms:
2 (h L , h R ) =

K
[h L (k) h R (k)]2
1
2 k=1 h L (k) + h R (k)

We repeat the test with the histograms of D and

D0 D+ and retain the maximum of the two result2
ing values, which we denote LR
. We can convert this

Figure 9. Illustration of half windows used for the estimation of

the texturedness. The texturedness of a label is based on a 2 test
on the textons in the two sides of a box as shown above for two
sample pixels. The size and orientation of the box is determined by
the selected scale and dominant orientation for the pixel at center.
Within the rocky area, the texton statistics are very similar, leading
to a low 2 value. On the edge of the wing, the 2 value is relatively
high due to the dissimilarity of the textons that fire on either side of
a step edge. Since in the case of the contour the contour itself can
lie along the diameter of the circle, we consider two half-window
partitions: one where the thin strip around the diameter is assigned
to the left side, and one where it is assigned to the other. We consider
both possibilities and retain the maximum of the two resulting 2
values.

to a probability-like value using a sigmoid as follows:

ptexture = 1

1
  2

1 + exp LR
/

(5)

This value, which ranges between 0 and 1, is small if

the distributions on the two sides are very different and
large otherwise. Note that in the case of untextured regions, such as a brightness step edge, the textons lying
along and parallel to the boundary make the statistics
of the two sides different. This is illustrated in Fig. 9.
Roughly, ptexture 1 for oriented energy maxima in texture and ptexture 0 for contours. ptexture is defined to
be 0 at pixels which are not oriented energy maxima.
4.3.2. Gating the Contour Cue. The contour cue is
gated by means of suppressing contour energy according to the value of ptexture . The gated value, p B , is defined as
p B = (1 ptexture ) pcon

(6)

In principle, this value can be computed and dealt with

independently at each filter scale. For our purposes, we
found it sufficient simply to keep the maximum value

Contour and Texture Analysis

19

Figure 10. Gating the contour cue. Left: original image. Top: oriented energy after nonmaximal suppression, OE . Bottom: 1 ptexture . Right:
p B , the product of 1 ptexture and pcon = 1 exp(OE /IC ). Note that this can be thought of as a soft edge detector which has been
modified to no longer fire on texture regions.

Figure 11. Gating the texture cue. Left: original image. Top: Textons label, shown in pseudocolor. Middle: local scale estimate (i). Bottom:
1 ptexture . Darker grayscale indicates larger values. Right: Local texton histograms at scale (i) are gated using ptexture as explained in 4.3.3.

of p B with respect to . The gated contour energy is

illustrated in Fig. 10, right. The corresponding weight
is then given by
WijIC = 1 max p B (x)
xMij

4.3.3. Gating the Texture Cue. The texture cue is

gated by computing a texton histogram at each pixel
which takes into account the texturedness measure
ptexture (see Fig. 11). Let h i be the K -bin texton histogram computed using Eq. (2). We define a (K + 1)bin histogram h i by introducing a 0th bin. The intuition

is that the 0th bin will keep a count of the number of

pixels which do not correspond to texture. These pixels arise in two forms: (1) pixels which are not oriented
energy maxima; (2) pixels which are oriented energy
maxima, but correspond to boundaries between two regions, thus should not take part in texture processing to
avoid the problems discussed in (1). More precisely,
h i is defined as follows:

ptexture ( j) I [T ( j) = k] k = 1 . . . K
h i (k) =
jN (i)

h i (0) = N B +

(1 ptexture ( j))

jN (i)

20

Malik et al.

where N (i) denotes all the oriented energy maxima

lying inside the window W(i) and N B is the number
of pixels which are not oriented energy maxima.
4.3.4. Combining the Weights. After each cue has
been gated by the above procedure, we are free to perform simple multiplication of the weights. More specifically, we first obtain W IC using Eq. (6). Then we obtain
W TX using Eq. (4) with the gated versions of the histograms. Then we simply define the combined weight
as
Wij = WijIC WijTX
4.3.5. Implementation Details. The weight matrix is
defined between any pair of pixels i and j. Naively, one
might connect every pair of pixels in the image. However, this is not necessary. Pixels very far away from
the image have very small likelihood of belonging to
the same region. Moreover, dense connectivity means
that we need to solve for the eigenvectors of a matrix
of size Npix Npix , where Npix is close to a million for
a typical image. In practice, a sparse and short-ranged
connection pattern does a very good job. In our experiments, all the images are of size 128 192. Each
pixel is connected to pixels within a radius of 30. Furthermore, a sparse sampling is implemented such that
the number of connections is approximately constant
at each radius. The number of non-zero connections
per pixel is 1000 in our experiments. For images of
different sizes, the connection radius can be scaled appropriately.
The parameters for the various formulae are given
here:
1. The image brightness lies in the range [0, 1].
2. IC = 0.02 (Eq. (1)).
3. The number of textons computed using K -means:
K = 36.
4. The textons are computed following a contrast normalization step, motivated by Webers law. Let
|F(x)| be the L 2 norm of the filter responses at
pixel x. We normalize the filter responses by the
following equation:

F(x) F(x)

log 1 +

|F(x)|
0.03

|F(x)|

5. TX = 0.025 (Eq. (4)).

6. = 0.3 and = 0.04 (Eq. (5))

Note that these parameters are the same for all the results shown in (6).
5.

Computing the Segmentation

With a properly defined weight matrix, the normalized cut formulation discussed in (3) can be used to
compute the segmentation. However, the weight matrix defined in the previous section is computed using
only local information, and is thus not perfect. The
ideal weight should be computed in such a way that
region boundaries are respected. More precisely, (1)
texton histograms should be collected from pixels in a
window residing exclusively in one and only one region. If instead, an isotropic window is used, pixels
near a texture boundary will have a histogram computed from textons in both regions, thus polluting
the histogram. (2) Intervening contours should only be
considered at region boundaries. Any responses to the
filters inside a region are either caused by texture or are
simply mistakes. However, these two criteria mean that
we need a segmentation of the image, which is exactly
the reason why we compute the weights in the first
place! This chicken-and-egg problem suggests an iterative framework for computing the segmentation. First,
use the local estimation of the weights to compute a segmentation. This segmentation is done so that no region
boundaries are missed, i.e. it is an over-segmentation.
Next, use this intial segmentation to update the weights.
Since the initial segmentation does not miss any region
boundaries, we can coarsen the graph by merging all
the nodes inside a region into one super-node. We can
then use these super-nodes to define a much simpler
segmentation problem. Of course, we can continue this
iteration several times. However, we elect to stop after
1 iteration.
The procedure consists of the following 4 steps:
1. Compute an initial segmentation from the locally
estimated weight matrix.
2. Update the weights using the initial segmentation.
3. Coarsen the graph with the updated weights to reduce the segmentation to a much simpler problem.
4. Compute a final segmentation using the coarsened
graph.
5.1.

Computing the Initial Segmentation

Computing a segmentation of the image amounts

to computing the eigenvectors of the generalized

Contour and Texture Analysis

eigensystem: (DW)v = Dv (Eq. (3)). The eigenvectors can be thought of as a transformation of the image
into a new feature vector space. In other words, each
pixel in the original image is now represented by a vector with the components coming from the corresponding pixel across the different eigenvectors. Finding a
partition of the image is done by finding the clusters in
this eigenvector representation. This is a much simpler
problem because the eigenvectors have essentially put
regions of coherent descriptors according to our cue
of texture and contour into very tight clusters. Simple
techniques such as K -means can do a very good job
in finding these clusters. The following procedure is
taken:
1. Compute the eigenvectors corresponding to the second smallest to the twelfth smallest eigenvalues of
the generalized eigensystem ((D W)v = Dv).6
Call these 11 eigenvectors vi , i = 2, . . . , 12. The
corresponding eigenvalues are i , i = 2, . . . , 12.
2. Weight7 the eigenvectors according to the eigenvalues: vi = 1 vi , i = 2, . . . , 12. The eigenvali
ues indicate the goodness of the corresponding
eigenvectors. Now each pixel is transformed to an
11 dimensional vector represented by the weighted
eigenvectors.
3. Perform vector quantization on the 11 eigenvectors
using K -means. Start with K = 30 centers. Let the
corresponding RMS error for the quantization be
e . Greedily delete one center at a time such that
the increase in quantization error is the smallest.
Continue this process until we arrive at K centers
when the error e is just greater than 1.1 e .
This partitioning strategy provides us with an initial
segmentation of the image. This is usually an oversegmentation. The main goal here is simply to provide
an initial guess for us to modify the weights. Call this
initial segmentation of the image S0 . Let the number of
segments be N0 . A typical number for N0 is 10100.

Figure 12.

p B is allowed to be non-zero only at pixels marked.

21

It should be noted that this strategy for using multiple eigenvectors to provide an initial oversegmentation
is merely one of a set of possibilities. Alternatives include recursive splitting using the second eigenvector
or first converting the eigenvectors into binary valued
vectors and using those simultaneously as in Shi and
Malik (2000). Yet another hybrid strategy is suggested
in Weiss (1999). We hope that improved theoretical insight into spectral graph partitioning will give us a better way to make this, presently somewhat ad hoc choice.
5.2.

Updating Weights

The initial segmentation S0 found in the previous step

can provide a good approximation to modify the weight
as we have discussed earlier. With S0 , we modify the
weight matrix as follows:
To compute the texton histograms for a pixel in Rk ,
textons are collected only from the intersection of
Rk and the isotropic window of size determined by
the scale, .
p B is set to zero for pixels that are not in the region
boundaries of S0 .
The modified weight matrix is an improvement over
the original local estimation of weights.
5.3.

Coarsening the Graph

By hypothesis, since S0 is an over-segmentation of the

image, there are no boundaries missed. We do not need
to recompute a segmentation for the original problem
of N pixels. We can coarsen the graph, where each
node of the new graph is a segment in S0 . The weight
between two nodes in this new graph is computed as
follows:

W kl =
Wij
(7)
iRk jRl

22

Malik et al.

Figure 13.

Initial segmentation of the image used for coarsening the graph and computing final segmentation.

Figure 14.

Segmentation of images with animals.

Contour and Texture Analysis

Figure 15.

Segmentation of images with people.

where Rk and Rl indicate segments in S0 (k and

l {1, . . . , N0 }); W is the weight matrix of the coarsened graph and W is the weight matrix of the original graph. This coarsening strategy is just an instance
of graph contraction (Chung, 1997). Now, we have
reduced the original segmentation problem with an
N N weight matrix to a much simpler and faster
segmentation problem of N0 N0 without losing in
performance.

5.4.

23

Computing the Final Segmentation

After coarsening the graph, we have turned the segmentation problem into a very simple graph partitioning

problem of very small size. We compute the final segmentation using the following procedure:
1. Compute the second smallest eigenvector for the
generalized eigensystem using W .
2. Threshold the eigenvector to produce a bipartitioning of the image. 30 different values uniformly spaced within the range of the eigenvector
are tried as the threshold. The one producing a partition which minimizes the normalized cut value is
chosen. The corresponding partition is the best way
to segment the image into two regions.
3. Recursively repeat steps 1 and 2 for each of the
partitions until the normalized cut value is larger
than 0.1.

24

Malik et al.

Figure 16.

5.5.

Segmentation of images of natural and man-made scenes.

Segmentation in Windows

The above procedure performs very well in images with

a small number of groups. However, in complicated
images, smaller regions can be missed. This problem
is intrinsic for global segmentation techniques, where
the goal is find a big-picture interpretation of the image.
This problem can be dealt with very easily by performing the segmentation in windows.
Consider the case of breaking up the image into
quadrants. Define Q i to be the set of pixels in the ith
4
Q i = Image. Exquadrant. Q i Q j = and i=1
tend each quadrant by including all the pixels which
are less than a distance r from any pixels in Q i , with r
being the maximum texture scale, (i), over the whole

image. Call these enlarged windows Q i . Note that these

windows now overlap each other.
Corresponding to each Q i , a weight matrix W i is
defined by pulling out from the original weight matrix
W the edges whose end-points are nodes in Q i . For each
W i , an initial segmentation S i0 is obtained, according
to the procedure in (5.1). The weights are updated as
in (5.2). The extension of each quadrant makes sure
that the arbitrary boundaries created by the windowing
do not affect this procedure:
Texton histogram upgrade For each pixel in Q i , the
largest possible histogram window (a (2 +1)2 box)
is entirely contained in Q i by virtue of the extension.

Contour and Texture Analysis

Figure 17.

25

Segmentation of paintings.

This means the texton histograms are computed from

all the relevant pixels.
Contour upgrade The boundaries in Q i are a proper
subset of the boundaries in Q i . So, we can set the
values of p B at a pixel in Q i to be zero if it lies on
a region boundary in Q i . This enables the correct
computation of WijIC . Two example contour update
maps are shown in Fig. 12.
Initial segmentations can be computed for each Q i
to give S i0 . They are restricted to Q i to produce Si0 .
These segmentations are merged to form an initial seg4
mentation S0 = i=1
Si0 . At this stage, fake boundaries
from the windowing effect can occur. Two examples
are shown in Fig. 13. The graph is then coarsened and

the final segmentation is computed as in (5.3) and

(5.4).
6.

Results

We have run our algorithm on a variety of natural images. Figures 1417 show typical segmentation results.
In all the cases, the regions are cleanly separated from
each other using combined texture and contour cues.
Notice that for all these images, a single set of parameters are used. Color is not used in any of these examples and can readily be included to further improve
the performance of our algorithm.8 Figure 14 shows
results for animal images. Results for images containing people are shown in Fig. 15 while natural and

26

Malik et al.

man-made scenes appear in Fig. 16. Segmentation results for paintings are shown in Fig. 17. A set of
more than 1000 images from the commercially available Corel Stock Photos database have been segmented
using our algorithm.9
Evaluating the results against ground truthWhat
is the correct segmentation of the image?is a challenging problem. This is because there may not be a
single correct segmentation and segmentations can be
to varying levels of granularity. We do not address this
problem here; a start has been made in recent work in
our group (Martin et al., 2000).
Computing times for a C++ implementation of the
entire system are under two minutes for images of size
108176 pixels on a 750 MHz Pentium III machine.
There is some variability from one image to another
because the eigensolver can take more or less time to
converge depending on the image.
7.

3.
4.
5.
6.
7.

8.

9.

necessary. Merging in this manner decreases the number of channels needed but necessitates the use of phase-shift information.
This is set to 3% of the image dimension in our experiments. This
is tied to the intermediate scale of the filters in the filter set.
This is set to 10% of the image dimension in our experiments.
Finding the true optimal partition is an NP-hard problem.
The eigenvector corresponding to the smallest eigenvalue is constant, thus useless.
Since normalized cut can be interpreted as a spring-mass system
(Shi and Malik, 2000), this normalization comes from the equipartition theorem in classical statistical mechanics which states that
if a system is in equilibrium, then it has equal energy in each mode
(Belongie and Malik, 1998).
When color information is available, the similarity Wij becomes
a product of 3 terms: Wij = WijIC WijTX WijCOLOR . Color similarity, WijCOLOR , is computed using 2 differences over color
histograms, similar to texture measured using texture histograms.
Moreover, color can clustered into colorons, analogous to textons.
These results are available at the following web page: http://
www.cs.berkeley.edu/projects/vision/Grouping/overview.html

Conclusion
References

In this paper we have developed a general algorithm

for partitioning grayscale images into disjoint regions
of coherent brightness and texture. The novel contribution of the work is in cue integration for image
segmentationthe cues of contour and texture differences are exploited simultaneously. We regard the experimental results as promising and hope that the paper
will spark renewed research activity in image segmentation, one of the central problems of computer vision.
Acknowledgments
The authors would like to thank the Berkeley vision
group, especially Chad Carson, Alyosha Efros, David
Forsyth, and Yair Weiss for useful discussions during
the development of the algorithm. We thank Doron
Tal for implementing the algorithm in C++. This research was supported by (ARO) DAAH04-96-1-0341,
the Digital Library Grant IRI-9411334, NSF Graduate
Fellowships to SB and JS and a Berkeley Fellowship
to TL.
Notes
1. For more discussions and variations of the K -means algorithm,
the reader is referred to Duda and Hart (1973) and Gersho and
Gray (1992).
2. It is straightforward to develop a method for merging translated
versions of the same basic texton, though we have not found it

Belongie, S., Carson, C., Greenspan, H., and Malik, J. 1998. Colorand texture-based image segmentation using EM and its application to content-based image retrieval. In Proc. 6th Int. Conf.
Computer Vision, Bombay, India, pp. 675682.
Belongie, S. and Malik, J. 1998. Finding boundaries in natural images: A new method using point descriptors and area completion.
In Proc. 5th Euro. Conf. Computer Vision, Freiburg, Germany, pp.
751766.
Binford, T. 1981. Inferring surfaces from images. Artificial Intelligence, 17(13):205244.
Canny, J. 1986. A computational approach to edge detection. IEEE
Trans. Pat. Anal. Mach. Intell., 8(6):679698.
Chung, F. 1997. Spectral Graph Theory, AMS. Providence, RI.
DeValois, R. and DeValois, K. 1988. Spatial Vision. Oxford
University Press. New York, N.Y.
Duda, R. and Hart, P. 1973. Pattern Classification and Scene Analysis, John Wiley & Sons. New York, N.Y.
Elder, J. and Zucker, S. 1996. Computing contour closures. In
Proc. Euro. Conf. Computer Vision, Vol. I, Cambridge, England,
pp. 399412.
Fogel, I. and Sagi, D. 1989. Gabor filters as texture discriminator.
Biological Cybernetics, 61:103113.
Geman, S. and Geman, D. 1984. Stochastic relaxation, Gibbs distribution, and the Bayesian retoration of images. IEEE Trans. Pattern
Anal. Mach. Intell., 6:721741.
Gersho, A. and Gray, R. 1992. Vector Quantization and Signal Compression, Kluwer Academic Publishers, Boston, MA.
Heeger, D.J. and Bergen, J.R. 1995. Pyramid-based texture analysis/synthesis. In Proceedings of SIGGRAPH 95, pp. 229238.
Jacobs, D. 1996. Robust and efficient detection of salient convex
groups. IEEE Trans. Pattern Anal. Mach. Intell., 18(1):2337.
Jones, D. and Malik, J. 1992. Computational framework to determining stereo correspondence from a set of linear spatial filters.
Image and Vision Computing, 10(10):699708.

Contour and Texture Analysis

Julesz, B. 1981. Textons, the elements of texture perception, and their

interactions. Nature, 290(5802):9197.
Knutsson, H. and Granlund, G. 1983. Texture analysis using twodimensional quadrature filters. In Workshop on Computer Architecture for Pattern Analysis and Image Database Management,
pp. 206213.
Koenderink, J. and van Doorn, A. 1987. Representation of local geometry in the visual system. Biological Cybernetics, 55(6):367
375.
Koenderink, J. and van Doorn, A. 1988. Operational significance of
receptive field assemblies. Biological Cybernetics, 58:163171.
Leung, T. and Malik, J. 1998. Contour continuity in region-based
image segmentation. In Proc. Euro. Conf. Computer Vision, Vol. 1,
H. Burkhardt and B. Neumann (Eds.). Freiburg, Germany, pp. 544
559.
Leung, T. and Malik, J. 1999. Recognizing surfaces using threedimensional textons. In Proc. Int. Conf. Computer Vision, Corfu,
Greece, pp. 10101017.
Malik, J., Belongie, S., Shi, J., and Leung, T. 1999. Textons, contours
and regions: Cue integration in image segmentation. In Proc. IEEE
Intl. Conf. Computer Vision, Vol. 2, Corfu, Greece, pp. 918925.
Malik, J. and Perona, P. 1990. Preattentive texture discrimination with
early vision mechanisms. J. Optical Society of America, 7(2):923
932.
Malik, J. and Perona, P. 1992. Finding boundaries in images. In Neural Networks for Perception, Vol. 1, H. Wechsler (Ed.). Academic
Press, pp. 315344.
Martin, D., Fowlkes, C., Tal, D., and Malik, J. 2000. A database of
human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics.
Technical Report UCB CSD-01-1133, University of California
at Berkeley. http://http.cs.berkeley.edu/projects/vision/Grouping/
overview.html.
McLean, G. 1993. Vector quantization for texture classification.
IEEE Transactions on Systems, Man, and Cybernetics, 23(3):637
649.
Montanari, U. 1971. On the optimal detection of curves in noisy
pictures. Comm. Ass. Comput., 14:335345.
Morrone, M. and Burr, D. 1988. Feature detection in human vision:
A phase dependent energy model. Proc. R. Soc. Lond. B, 235:221
245.

27

Morrone, M. and Owens, R. 1987. Feature detection from local energy. Pattern Recognition Letters, 6:303313.
Mumford, D. and Shah, J. 1989. Optimal approximations by piecewise smooth functions, and associated variational problems.
Comm. Pure Math., 42:577684.
Parent, P. and Zucker, S. 1989. Trace inference, curvature consistency, and curve detection. IEEE Trans. Pattern Anal. Mach. Intell., 11(8):823839.
Perona, P. and Malik, J. 1990. Detecting and localizing edges composed of steps, peaks and roofs. In Proc. 3rd Int. Conf. Computer
Vision, Osaka, Japan, pp. 5257.
Puzicha, J., Hofmann, T., and Buhmann, J. 1997. Non-parametric
similarity measures for unsupervised texture segmentation and
image retrieval. In Proc. IEEE Conf. Computer Vision and Pattern
Recognition, San Juan, Puerto Rico, pp. 267272.
Raghu, P., Poongodi, R., and Yegnanarayana, B. 1997. Unsupervised
texture classification using vector quantization and deterministic
relaxation neural network. IEEE Transactions on Image Processing, 6(10):13761387.
Shaashua, A. and Ullman, S. 1988. Structural saliency: The detection of globally salient structures using a locally connected network. In Proc. 2nd Int. Conf. Computer Vision, Tampa, FL, USA,
pp. 321327.
Shi, J. and Malik, J. 1997. Normalized cuts and image segmentation.
In Proc. IEEE Conf. Computer Vision and Pattern Recognition,
San Juan, Puerto Rico, pp. 731737.
Shi, J. and Malik, J. 2000. Normalized cuts and image segmentation.
IEEE Trans. Pattern Anal. Mach. Intell., 22(8):888905.
Weiss, Y. 1999. Segmentation using eigenvectors: A unifying view.
In Proc. IEEE Intl. Conf. Computer Vision, Vol. 2, Corfu, Greece,
pp. 975982.
Wertheimer, M. 1938. Laws of organization in perceptual forms (partial translation). In A Sourcebook of Gestalt Psychology, W. Ellis
(Ed.). Harcourt Brace and Company, pp. 7188.
Williams, L. and Jacobs, D. 1995. Stochastic completion fields: A
neural model of illusory contour shape and salience. In Proc. 5th
Int. Conf. Computer Vision, Cambridge, MA, pp. 408415.
Young, R.A. 1985. The Gaussian derivative theory of spatial vision: Analysis of cortical cell receptive field lineweighting profiles. Technical Report GMR-4920, General Motors
Research.