The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
1218
band width hi of x¡. Now, we can get a practical algorithm for
mean shift iteration.
Let Xi and n be the «/-dimensional input and filtered result of x h
n is the number of pixels. Mean Shift based image segmentation
algorithm can be described as follows:
For i = l:n
Estimate h t ;
End
For i = l:n
Initialize j=l and y i 1 = X■;
Compute y i /+1 iterative using (12) until convergence;
z i=y iJ+ 1’
End
K-means for finally segmentation result
3.2 Mean Shift Segmentation Implementation in High
Dimension Feature Space
In our method, original fixed-bandwidth Mean Shift algorithm
is replaced by a variable-bandwidth technology in (Georgescu
et al.) which is suitable for the segmentation task in high-
dimension features.
Spectral feature in Lab color space with 3 dimensions and PCs
got from Gabor features corresponding to different stages and
orientations of image frequency domain are combined together
to consist of feature space. However, even about a half of PCs
of Gabor features is reduced; there are about 35 Gabor features
should be preserved to maintain the balance between
performance and processing time for a 3 bands multispectral
Quickbird image using a Gabor filter banks with 4 stages and 6
orientations.
In practise, the Mean Shift iteration of (12) require a
neighbourhood query, it is obviously impractical by scanning
the whole feature space in such high dimension space,
considering the high complexity of the mean shift algorithm
0(Ndn 2 ),where N is a average iteration times for one point. In
order to improve the efficiency of the neighbourhood query, a
data structure named locality- sensitive hashing(LSH) is used in
(Georgescu et al.). Limited by the space, here we only give a
brief review on it.
The basic idea of LSH is to hash the input data so that similar
items are mapped to the same buckets with high probability.
The algorithm has two main parameters: the width parameter k
and the number of hash tables L. Firstly, a family of hash
functions H is defined; and then L group hash functions g is
obtained by concatenating k randomly chosen hash functions L
times from set H. In other words, the algorithm constructs L
hash tables, each corresponding to a different randomly chosen
hash function g and g is consisted with k randomly chosen hash
functions in H.
Figured The flow chart of our segmentation technology
Now, a segmentation flow chart is given. Firstly, HR multi
spectral image with RGB bands is filtered by Gabor filter banks
respectively and the dimension of spatial features is reduced by
PCA. On another way, 3-dimensional spectral features obtained
by converting pixel values in RGB color space to Lab. It is
evident that the measure unit of feature has great influence on
the clustering, The usual form of equalizing the feature
contribution consists of performing some scaling
operation(Sa,200l),section2.3.However, it is almost impossible
to devise a suitable standardization method , since we don’t
know beforehand the type of the clusters we are dealing with.
One feasible way is “trial and error” approach. In our
experiment, each dimension of Gabor features is normalized by
scaling into the range [0 1], and scale Lab color space data by
multiplying a constant 4/125. After pre-processing of image
feature, the adaptive Bandwidth Mean Shift smoothing
portrayed in the secition3.1 is executed in the joint feature
space. Finally, segmentation results are analysized.
4. EXPERIMENTS RESULTS
4.1 Experiment Images and Algorithm Parameters
In this part, some experiments have been conducted to illustrate
the effectiveness of the proposed method. In the experiments,
test images with RGB three bands are cropped from a
QuickBird imagery of Wuhan, China. As shown in follow
Figure 2, the first image consists of three typical land cover
types: naked land, grass and tree. It is obvious that the tree is
very close with grass in spectral information. The second image
consists of five classes, that is, water body, farm land, tree,
building and shadow.
As mentioned in Section 3, there are 7 parameters used in the
proposed segmentation method. In Gabor filter banks design
stage, the centre frequency rang of interest and the scale and
orientations parameter are needed. In our experiment, we found
that the default set of parameters in (Manjunath and Ma,1996)
can meet ore demand on efficiency and performance. [0.04 0.5]
is set to frequency center rang and 4 scales and 6 directions is
chosen. For PCA, the contribution threshold used for preserving
principle components is 0.98. In hashing the high dimension
set stage, L and K parameters described in section 3.2 for the
first image are 2 and 28. Finally, the number of class in image,
the first is 3, the second is five.