1224
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
The paper is organized as follows. In section one some
background and the objective of this research are introduced.
Section two introduces the feature extraction method in this
paper. The segmentation methodology is presented in section
three. In section four, we carry out several experiments and
demonstrate the segmentation results. Section five concludes
the paper with discussion and conclusion.
2. FEATURE DESCRIPTION
In the whole segmentation process, we utilize a novel texture
and spectral feature extraction method which considers the
cross-band relations between pixels. Principal Component
Analysis (PCA) is adopted in this study to get rid of redundant
information and make it convenient to extract texture features
of multispectral images. More specifically, we obtain the first
two principal components (PCs) of multispectral (i.e. blue,
green, red and near infrared) H-res satellite imagery through
PCA. They collect most of the information of the H-res
imagery. Then the texture and spectral information is calculated
from the transformed PCs. The distributions of texture and
spectral information are denoted by discrete two-dimensional
histograms whose two dimensions correspond to the two PC
variables respectively.
2.1 LBP texture operator
The texture analysis operator of LBP (Local Binary Pattern)
was first introduced as a complementary measure of local image
contrast by Ojala et al. (Ojala et al. 1996) and was extended by
subsequent studies. Due to its major advantages on simple
theory, computational simplicity and robustness to rotation and
monotonic transformation of gray scale, it has been frequently
used in many studies, such as texture segmentation or
classification (Chen and Chen 2002; Ojala et al. 2006), moving
objects detection (Heikkila and Pietikainen 2006) and
segmentation of remote sensing imagery (Hu et al. 2005;
Lucieer et al. 2005).
The name “Local Binary Pattern” reflects the functionality of
the operator, i.e., a local neighborhood is thresholded at the
gray value of the center pixel into a binary pattern (Ojala et al.
2002). The original LBP was produced by multiplying the
thresholded values with weights given to the corresponding
pixels, and summing up the results (Maenpaa 2003). Ojala et al.
(Ojala et al. 2002) proposed gray-scale invariance LBP
form LBP p r , which is defined as
P-1
LBP 2'
(1)
P=0
Where
s „= s (g P -g c ) =
o, g P < g.
1> g D ^g c
(2)
Where P is the number of neighboring pixels on a circle of
radius R, g c corresponds to the gray value of the centre pixel
of a texture unit and g p is the gray value of its neighbourhood.
In order to achieve rotation invariance, Ojala et al. (Ojala et al.
2002) presented the term of ‘Uniform’, whose measure
corresponds to the number of spatial transitions (bitwise 0/1
changes) in the patterns. However, the ‘uniform’ pattern is
defined in the case of regular textures i.e. Brodatz’s textures,
which consist of the vast majority of “uniform” patterns of all 3
x 3 patterns and it is not the case of satellite imagery through
our experiment. So we present a rotation invariant LBP form
LBPp R that is more suitable for describing natural scenes:
LBPj„ = IX ! ROR(LBP c , i) I i = 0, 1, • • •, P-1}
P i=0
(3)
Where the ROR is defined as:
ROR(LBP p R ,i) =
(4>
LBP,
P,R »
i = 0
2.2 Texture and spectral distribution
The texture feature is extracted on gray level images in the most
of the previous studies. For multi-spectral imagery, it does not
consider cross-band relations (Hu et al. 2005). Although
Lucieer et al. (Lucieer et al. 2005) considered the cross-band
relations by multivariate texture model, the method is too
complicated. In this paper, the texture feature of an image
region is evaluated by the joint distribution, i.e. a discrete two-
dimensional histogram, of LBP operator operated on two PCs of
the image region. In the following experiments, we apply
LBP S " to calculate the texture distribution LBP ] / LBP\ of an
image region and compare their efficiency in colour image
segmentation. The spectral feature of an image region is just the
joint distribution of grayscale values of its two PCs. As the
number of bins used in the quantization of the feature space is a
trade-off between the discriminative power and the stability of
the feature transform, we set the bins of spectral distribution as
32 by 32 in the following study.
2.3 Similarity measure
In the split and merge segmentation process, we choose a non-
parametric statistic the G-statistic as a pseudo-metric for
comparing the similarity between texture and spectral
distributions. The similarity between a sample and model
histograms is computed by the formula:
By our experiments, In the split and merge segmentation
process, we choose a non-parametric statistic the G-statistic as a
pseudo-metric for comparing the similarity between two
histograms and the similarity between two regions i and j is
measured by weighted sum G-statistic WG(i,j) of the
similarity measures of three features. Then the similarity
between two regions i and j is measured by weighted sum G-