The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
that the latter one overrates the size of regions and makes the
segmentation results unstable. In the merging process, we
utilize RAG (Region Adjacency Graph) to describe the blocks
after splitting. The RAG consists of three components: V, E
and M. V is a set of region nodes to record the region
information. E is an adjacency matrix to record the pseudo
address of each region edge. The M matrix records the merger
importance values of all the pairs of adjacent regions of an
image. At each merging step, we search the smallest MI of the
M matrix and merge the pair of adjacent regions that has the
smallest MI. Then we adjust the RAG, and merge the pair of
adjacent regions that has the smallest MI in the changed M
matrix. Merging proceeds until the following stopping criteria
is true:
MIR = CJ ^>Y
Ml
max . (12)
Where MI cur and MI max denote the merger importance for the
current best merge and the largest merger importance of all
preceding mergers. Threshold Y is determined experimentally.
A boundary refinement algorithm is used to refine the
boundaries of the blocky segmented image. For an examined
boundary point P, a discrete square with a dimension d around
the pixel is placed and the MI between the square and the
neighboring regions of point P is computed. The pixel is
relabelled if the label of the neighboring region that has the
smallest MI is different from the label of P. At the following
step, we only consider the boundary points that have relabelled
at the previous sweep. The procedure is iterative and proceeds
until the un-relabelled number of the boundary pixels is less
than 50 or the iteration times are larger than 30.
4. EXPERIMENTS AND RESULTS
The objective of the present experiments was to evaluate the
effectiveness of the novel features of texture and spectral
distributions and the very simple weight combination approach
in segmentation of H-res remote sensing imagery. Besides,
we discuss the effect of several parameters, i.e. weight
determination, MI and thresholds, on the result of segmentation
for the purpose of obtaining improved results and finding a way
of solving SIMF better.
The performance of the method was evaluated with 256x256
pixel multi-spectral IKONOS-2 satellite images. IKONOS-2
data contain red, green, blue and near-infrared (NIR) channels
at 4.0 m spatial resolution. Since the colour images are the most
common in application and can provide more information than
grayscale images, the paper is endeavour to explore
segmentation approach that make good use of multi-spectral or
colour information.
The experiments were performed using the following
procedure. The original multi-spectral images are transformed
by PCA. We just take the first two PCs for feature extraction.
They collect more than 95% information of the original images.
For texture features, we computed texture labelled images of
the PCs by rotation invariant LBP form and we got two LBP
labelled images which were used to obtain the discrete two
dimensional texture histograms. The texture similarity of two
regions was calculated by their two-dimensional texture
histograms. The spectral histogram was gotten by their joint
distribution of the gray-scale pixel values of the PCs. So we got
the spectral similarity of two regions from their spectral
histograms. The first PC was used to calculate the attribute of
regions by their standard deviation, which was applied to
weight determination of the two features. The weighted sum
similarity measures were used to the whole coarse to fine
segmentation process.
Figure 2. Segmentation results of H-res images based on
texture distributions calculated by LBP^\ and
spectral distributions.
Figure 2 shows the segmentation results by SIMF approach
based on texture distributions calculated by LBP^\ and
spectral distributions. The result demonstrates that the SIMF
approach by our feature extraction method performs well on
complex H-res satellite images.
5. DISCUSSION AND CONCLUSION
The paper presented a novel feature extraction method that
considers the cross band relations and a new segmentation
framework SIMF suitable for segmenting multi-spectral images.
Figure 2 demonstrates the satisfied segmentation results. It
shows thatLBP%\ is a robust LBP operator for texture feature
extraction. Despite that, the feature weight determination is still
a necessary research topic in the future since the images are
very complex and different textures may be used to SIMF.
Based on the previous experiments and results, we point out the
future research works. The feature extraction method is very
important for SIMF. The future research should concentrate on
finding more appropriate features adaptive to different kinds of
images, e.g. that of various resolution. The feature weight
combination approach determines whether the combining
features can discriminate heterogeneous regions to a large
extent, which is still an open problem. The MI determines the
sequence of merging of pairs of homogeneous regions and the
stopping criterion for merging. MIR determines when to stop
the merging process and the scale of segmentation results. So
the future research should explore MIR that can implement
multiscale segmentation. Similarity measure of feature