1267
A REGION-BASED TECHNIQUE FOR FUSION OF HIGH RESOLUTION IMAGES
USING MEAN SHIFT SEGMENTATION
Li Shuang 3 , Li Zhilin 3 ’ b
a LIESMARS, Wuhan University, P.R. China - lishuangl29@gmail.com
b Dept, of Land Surveying and Geo-Informatics, The Hong Kong Polytechnic University, Hong Kong
- lszlli@polyu.edu.hk
Commission VII, WG VII/6
KEY WORDS: Image processing, Sharpening, Image understanding, Fusion, Geography
ABSTRACT:
This paper describes a region-based technique for fusion of high-resolution images. In this technique, mean shift segmentation is
adopted to extract the features for high resolution image as a substitution of other segmentation methods (e.g. Canny operator) and
Structure Similarity Index Metric (SSIM) is used to measure the region similarity. Experiments on IKONOS images are carried out
to compare the results obtained from this new technique and those with Canny segmentation. It has been found that the results
obtained this new technique is much better than the conventional ones in terms of spatial sharpness and spectral reservation.
1. INTRODUCTION
Image fusion is a process to combine two or more different
images to form a new image by using certain algorithm (Phol
and Genderen,1998). It takes place at three levels: pixel, feature
and decision. Image fusion at pixel-level is the lowest
processing level considering individual pixels or associated
local neighbourhoods of pixels for fusion decision. In the past
decades, large number of pixel level image fusion methods is
proposed, e.g. Brovey, Intensity-Hue-Saturation, Principle
Component Analysis, Wavelet transforms, etc (Zhen and
He,2004). However, such methods often introduce colour
distortion and/or block effect to high resolution image fusion.
This is because a pixel is only a basic unit of information with
no semantic significance. At the feature level, features from the
input images will be first extracted (e.g. using segmentation
procedures); and then fusion of these features will be operated
by some rules. Comparing with pixel level image fusion,
feature-level method is more meaningful. Because it can fully
explore the characteristics of features to guide the image fusion
process, such as region activity level, region similarity match
measures and so on.
More recently, a number of region-based feature-level image
fusion techniques have been proposed (Zhang, 1997; Piella,2003;
Lewis,2007). These techniques first transform the source
images A and B to multi-scale representations by wavelet
transforms; segmentation is carried out on the source image to
get region representations of both images. Then by overlaying
the two region representations, a shared region representation
for these two images is obtained. And region activity level and
similarity match measurements are calculated from each region
to guide the fusion process. During the whole process,
segmentation is the most important part because it directly
influences the effect of fusion result. Previous work employs
the watershed segmentation or Canny edge detection method
and the results for fusion of high-resolution images are not very
good. Therefore, this study aims to develop a new technique for
fusion of high-resolution images.
It has been found (e.g. Mo et al,2006) that the mean shift
segmentation is more suitable for the segmentation of high
resolution image and thus will be adopted in this study.
Moreover, the Structure Similarity Index Metric (SSIM)
proposed by Wang,(2002) for image quality assessment will be
used (instead of region match measure which is commonly used)
to guide the fusion process.
Section 2 reviews the region based fusion. Mean shift
segmentation for feature extraction is introduced in section 3.
SSIM used for fusion decision making is described in section 4.
Section 5 describes the evaluation of the proposed method and
conclusions are made in section 6.
2. REGION BASED FUSION: AN OVERVIEW AND A
PROPOSAL
The concept of region based fusion was first introduced by
Zhang et al.,(1997) and developed by Piella et al.,(2003).
Piella’s generic region based image fusion framework is shown
in Figure 1.