DETECTING TEXTURE EDGES FROM IMAGES
329
HE Dong-chen and WANG Li
Centre d'applications et de recherches en télédétection (CARTEL)
Université de Sherbrooke
Sherbrooke, (Québec)
CANADA J1K 2R1
ABSTRACT
Edge detection takes an important place in image processing
and pattern recognition. Its objective is to locate prominent
edges in an image, and so to separate the components of an
image into subsets that may correspond to the physical objects
in the scene. In general, this can be achieved by generating an
edge map using, for example, edge detection operators or
some threshold techniques. In most cases, these methods
assume that at the edge the grey level intensity changes in a
discontinuous way (usually as a step function). If we need to
segment a textural scene by finding the texture boundaries,
traditional methods of edge detection are usually not
successful sicne they cannot distinguish between the micro
edges within each texture and the boundaries between
different textures. One reason for their failure is their inability
to properly characterize a texture. This problem can be solved
by combining the traditional edge detection techniques with
some efficient textural measures. That is, in the edge detection
operators, grey levels are replaced by textural features.
Recently, a texture spectrum method has been proposed for
texture characterization. This paper presents an example of the
application of the texture spectrum to edge detection.
Promising results are obtained when locating texture
boundaries of some of Brodatz' natural images.
Key Words: Edge detection, Texture boundary, Texture
analysis, Texture spectrum.
1 INTRODUCTION
Texture is the term used to describe the surface of a given
phenomenon (the spatial intensity relationships between
pixels) in an image. Texture analysis plays an increasingly
important role in digital image processing and pattern
recognition, and is widely applied to the processing and
interpretation of remotely sensed data, and biomedical and
microscopic cell images, where texture information is
sometimes the only way to characterize a digital image.
One of the major researche works in texture analysis is to
determine the boundaries between the different textured
regions for a given image having several textured areas.
Traditional edge detection operators, such as Roberts, Sobel,
Prewitt, and Gaussian-Laplacian operators (Marr & Hildreth,
1980; Young & Fu, 1986), or some threshold techniques
(Haddon, 1988) are widely used in practice for extracting
prominent edges in an image. The basic assumption in these
approaches is that the intensity variation is more or less
constant within the region and takes a different value at its
boundary. Due probably to the lack of a formal definition of
texture and of texture boundaries, these traditional methods of
edge detection are not suitable for detecting texture boundaries
in an image. The major difficulty for this kind of approach is
in distiguishing between the texture boundaries which delimit
different textures, and the micro-edges located within a same
texture.
This difficulty may be overcome by combining the traditional
techniques of edge detection with texture features. That is, in
the edge detection operators, the grey levels of pixels are
replaced by textural measures. As texture features
characterize the texture aspects of an image, we may assume
that these features take more or less a constant value within a
same textured area, and take a different value at texture
boundaries. In this way, traditional edge detection can be
applied to extracting texture boundaries of images.
In the past years, a lot of features have been proposed to
extract the texture information of images. They can be broadly
divided into two categories: structural and statistical
approaches (Haralick, 1979; He et al. 1987; He et al. 1989).
In the structural approach, texture is considered as a repetition
or quasi-repetition of fundamental elements of images with a
certain rule of displacement, the Fourier spectrum analysis is
often used to extract the primitives and the displacement rule.
Texture features are then extracted from the Fourier spectrum
plan. In the statistical approach, the objective is to characterize
the stochastic properties of the spatial distribution of grey
levels in an image. Haralick's features derived from the co
occurrence matrix approach are widely used in practice
(Haralick et al., 1973). However in these methods, textures
are characterized by a set of features. Thus, using only one of
these features in edge detection operators risks limiting their
performance in the characterization of textures.
Recently, the texture spectrum method (He & Wang, 1989;
Wang & He, 1990) has been proposed as a new statistical
approach to texture analysis, where the local texture for a
given pixel and its neighbourhood is characterized by the
corresponding texture unit, and an image can be characterized
by its texture spectrum which is the occurrence frequency
function of all the texture units within the image. This method
has been used with success for texture classification (Wang &
He, 1989), textural filtering and for some geological
applications using remotely sensed data (Wang et al., 1989;
Wang & He, 1990). These experimental results show the
promising performance potential of the texture spectrum in the
characterization of textures.
In this paper, we present examples of the use of the texture
spectrum as a texture feature and applying it to texture edge
detection.
2 TEXTURE UNITS & TEXTURE SPECTRUM
This section gives a brief review of the texture unit and
texture spectrum. The method has been introduced and
described in detail elsewhere by He and Wang (He & Wang,
1989; Wang & He, 1990). The basic concept is that a texture
image can be considered as a set of essential small units
termed texture units, which characterize the local texture
information for a given pixel and its neighbourhood. The
statistics of all the texture units over the whole image reveal
the global texture aspects.
