Você está na página 1de 4

2008 International Conference on Intelligent Computation Technology and Automation

A new image fusion algorithm based on Fuzzy Logic

Zhu Mengyu Yang Yuliang


Department of Electronic Engineering, Department of Communication Engineering,
School of Information Science and School of Information engineering, University
Technology, Beijing Institute of Technology, of Science and Technology Beijing,Beijing,
Beijing, P.R. China, 100081 P.R. China, 100083
E-mail: zmy@bit.edu.cn E-mail: teacheryangustb@126.com

Abstract
2. The pixel-level image fusion
An image fusion algorithm is presented based on
fuzzy logic and wavelet in this paper. Aiming at the In the pixel-level image fusion method, the acquired
visible and infrared image fusion, we analyze the pixel- original image data are directly fused based on pixel by
level image fusion algorithms, and addresses an pixel. The pixel-level fusion can provide the detailed
algorithm based on the discrete wavelet transform and information that cannot be provided by the other levels
fuzzy logic. We created two fuzzy relations, and of fusion. And the results still are images, which are
estimated the importance of every wavelet coefficient not only more intuitive for human beings, but also
with fuzzy reasoning. According to the importance of more suitable for further processing.
coefficients, the weighting average coefficients were Vision Enhancement is the core task of the pixel-
calculated. Ultimately, we reconstructed the fused level image fusion technology. Multi-source image
image through inverse wavelet transform. In the fusion does not consider the physical characteristics of
experimental section, we verified the effectiveness of the original image. The purpose of the fusion is to fuse
the algorithm with two actual image data acquired the same scene of multi-source images into one image.
from the field experiments. The commonly used methods include the weighted
average method, selection and the weighted average
1. Introduction combined method.

The concept of Data Fusion was presented for 2.1 The algorithm combined by the selection
military fields in the 1970s [1]. In order to get the best and the weighted average method
operational results, the new weapon systems usually
use many sensors. The idea of data fusion is widely The selection and the weighted average combined
used in image processing. There are many fusion method, including two modes, i.e. the selection mode
approaches in the multi-source images fusion. The and the weighted average mode. For double-source
image fusion methods are divided into three levels: data fusion, this method must first determine which
pixel, feature and decision level. kind of mode to be used. The process of this method is
The main purpose of the pixel-level image fusion is as follows:
to enhance the quality of images. Some scholars have First, the transformation, such as the gradient
studied pixel-level image fusion issues, such as P.J. pyramid, the contrast pyramid, or wavelet transform,
Burt and R.J. Kolczynski [2] presented a weighted decomposes the original image data into several sub-
average algorithm based on the match and salience bands.
measures. Their algorithm improves the signal-to-noise Second, coefficients X and Y are chosen.
ratio, but deteriorates the contrast. Many scholars are Third, calculate the respective salience σ x and
carrying the further study on this algorithm.
In this paper, we studied the pixel-level images σy.
fusion of visible and infrared image, and a new Fourth, calculate the match coefficient. For
algorithm based on the wavelet and the fuzzy logic was example, it can be the covariance between X and Y.
presented.

978-0-7695-3357-5/08 $25.00 © 2008 IEEE 83


DOI 10.1109/ICICTA.2008.336
Fifth, if the match coefficient is larger than the professor Zadeh as an extension of the classical notion
threshold, one weighted coefficient Wmin is estimated, of set. Fuzzy set theory permits the gradual assessment
of the membership of elements in a set; this is
and another weighted coefficient is Wmax = 1 − Wmin . described with a membership function valued in the
Otherwise the selection mode is chosen, or real unit interval [0, 1]. Classical sets are special cases
Wmin = 0,Wmax = 1 . If σ x > σ y , Wx = Wmax , of the membership functions of fuzzy sets, if the latter
only take values 0 or 1.
W y = Wmin . Otherwise Wx = Wmin , W y = Wmax . The application of fuzzy logic is very broad, such as
Sixth, calculate the fused data with image processing, control, artificial intelligence and
other fields, and so on.
Z = Wx X + W yY . In our fusion algorithm, DB9/7 biorthogonal
The final step is an inverse transform using the wavelet [4] is used. We deal with four sub-bands with
processed data, and then the fused image is shown. two different rules. For approximate component, we
The basic idea of the algorithm is to implement the average them; for the horizontal, the diagonal and the
weighted average or selection in the transformation vertical detail components, we fuse them with our
domain, then get the fused image with reverse fuzzy algorithm.
transform. For every coefficient of the three detail components,
we must estimate their contribution to the fused image.
3. The fusion algorithm based on wavelet Consider if "a coefficient is large and the information
transform and fuzzy logic entropy of it is large", the coefficient may contribute
more importance to the fused image. We wish to
It is difficult to directly fuse multiple source images establish a fuzzy membership function for the
in the spatial domain. If the two-dimensional wavelet sentence. We begin by recognizing that this sentence is
is used to decompose the image, we can choose a a composition between two relations "a coefficient is
different fusion strategy for the different bands based large, " and "the information entropy of it is large, ".
on their different characteristics. Then we create membership functions for the relation
"a coefficient is large, " μ 0 ( x) , and "the information
3.1 The wavelet transform entropy of it is large, " μ1 ( x ) . Finally, using these
membership functions and fuzzy reasoning, we can
Wavelet transforms [3] are mathematical functions
create the membership function for sentence, as
that cut up data into different frequency components,
and then study each component with a resolution μ0∩1 ( x) = min( μ0 ( x), μ1 ( x)) , which represents the
matched to its scale. They have advantages over importance of the coefficient to the fused image.
traditional Fourier transforms in the situations where For a coefficient x in a given sub-band A of the first
the signal contains discontinuities and sharp spikes.
Image processing is the one of the main areas of image, we define a fuzzy membership function
μ 0 ( x)
wavelet transforms. of the first relation as,
We use two-dimensional DWT for image abs( x)
processing in this paper. Two-dimensional DWT μ 0 ( x) = (1)
decomposes a two-dimensional signal at different max(abs( A))
scales into the approximation and details of the original μ ( x) of the second
Fuzzy membership function 1
signal.
The results of two-dimensional decomposition are relation is defined as,
the raw data, the horizontal, the diagonal and the log 2 ( p( x))
μ1 ( x) = (2)
vertical detail. min (log 2 ( p( y )))
The wavelet coefficients are combined in the y∈A
transform domain using a defined fusion rule then where p (x ) means the probability of the
transformed back to the spatial domain to give the coefficient x in the sub-band A.
fused image. Then, we calculate the importance of the coefficient
with fuzzy reasoning as
3.2 Fuzzy fusion algorithm
μ 0∩1 ( x) = min(μ 0 ( x), μ1 ( x)) (3)
Fuzzy sets are sets whose elements have degrees of For the coefficient y, which has the same spatial
membership. Fuzzy sets have been conceived by position with the coefficient x, in the sub-band A of the

84
second image, we can also get the fuzzy membership
functions for two relations μ0 ( y) and
μ1 ( y ) .
We fuse three sub-band coefficients using every
coefficient’s μ 0∩1 ( x ) or μ 0∩1 ( y ) . The fused
coefficient z is calculated as
x ⋅ μ 0∩1 ( x) + y ⋅ μ 0∩1 ( y )
z= (4)
μ 0∩1 ( x) + μ 0∩1 ( y )
After every coefficient has been processed, the
inverse wavelet transform is applied to reconstruct the
fused image.
Figure 3. The fused image
4. The experiment and analysis
Figure 1 shows the original visible image. We can
In this section, we fused a visible image and an see a small plane, whose details are clear. Figure 2
infrared image using our algorithm. These images are shows the original infrared image. From this image, we
acquired from the field experiments. are easy to see the engine of the plane, which is not
notable in Figure 1, but the other details of this image
are poor. We processed two images with our fuzzy
logic algorithm. The fused image is displayed in Figure
3. We can see the windows of the plane, the engine, the
trunk of trees. In one word, this image provides more
information than anyone of the original images.
Table1. The evaluation of images fusion
mean std Gradient Spatial Entropy
frequency
Visible 144.3 46.8 5.6 13.7 7.6
IR 151.2 29.2 5.2 13.3 7.7
Fused 173.8 44.7 7.3 18.2 7.9
Some parameters have been calculated to evaluate
Figure1. The original visible image our fusion algorithm. They are shown in Table 1. The
result shows our fused image has largest gradient [5],
spatial frequency [6] and entropy. The objective
quality of the fused image is satisfied.

5. Conclusions
We proposed a pixel-level image fusion algorithm
based on fuzzy logic and wavelet in this paper. The
images are processed in wavelet transform domain. We
created two fuzzy relations, and estimated the
importance of every wavelet coefficient with fuzzy
reasoning. According to the importance of coefficients,
Figure 2. The original infrared image the weighting average coefficients were calculated.
Ultimately, the fused images are reconstructed through
inverse wavelet transform.
We also verified the effectiveness of the algorithm
through subjective and objective judgments with two
actual images.

6. References

85
[1] W. Edward, L. James, Multisensor data fusion, Artech [4] I. Daubechies, "The wavelet transforms, time-frequency
House, Boston, 1990. localization and signal analysis", IEEE Trans on Information
Theory, 1990, vol. 36, no. 5, pp. 961-1005.
[2] P.J.Burt, R.J.Kolczynski, "Enhanced image capture
though fusion", Proceedings of 4th International Conference [5] J.H. Park, K.O. Kim, Y.K. Yang, "Image Fusion Using
on Computer Vision, 1973, pp. 173-182 Multiresolution Analysis", Geosciences and Remote Sensing
Symposium, 2001, vol. 2, pp. 864-866.
[3] S. Mallat, "A theory for multi2resolution signal
decomposition: The wavelet representation", IEEE Trans on [6] A. Aldroubi, P. Abry, M. Unser, "Construction of
Patterns Analysis and Machine Intelligence, 1989, vol. 11, no. Biorthogonal Wavelets Starting from any two
7, pp. 674-693. Multiresolutions", IEEE Trans on Signal Processing,1998,
vol. 46, no. 4, pp. 1130-1133

86

Você também pode gostar