The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B6b. Beijing 2008 
215 
Figure 3 is processed image by trapezium low pass filter (TLPF). 
Results suggests: the noise will be removed when Dl=50. 
However, the ring is more serious when DO is bigger while the 
image is clearer. After processing, the image is blured than that 
before processing. 
Low pass filter can remove stripped noise well; however, much 
detailed information is also removed. So it is not fit to the 
hyperspectral image which needs much spectral information. 
4.2 Gray Value Substitution 
Because this noise is symmetrical and regular, we developed 
one simple method to reduce this stripped noise without 
Fourier transformation and wavelet transformation. This 
method firstly detects the noise value, then a value nearby is 
used to replace the noise value. After that, a median filter is 
used to do noise reduction further more. The result image can 
be seen in figure 3. 
Figure 4 Image processing results of gray value substition 
We can see from this image, most noise line is modified and the 
line without noise is not affected by this method. So the image 
reserves much intrinsic information. However, some small 
strippes still exsits, and the brightness after processing is larger 
than that before processing. 
4.3 Wavelet transformation 
Wavelet image denosing has been well acknowleged as an 
important method of image denoising. This method reserves 
most wavlet coefficient that contains information, so it can 
preserve image detail. Image denoising by wavelet uaually 
mainly has three steps: 
(1) Decomposition of image signal; 
Choose appropriate wavelet and right decomposition level (note 
N), then N levels decompostion and computation to 2D image 
signal. 
(2) Threshold quantification of high-frequency coefficient after 
level decomposition; 
Choose an appropriate threshold to each decompositon level, 
and soft threshold quantification to high-frequency coefficient. 
The choose rule of threshold is equal to the prior part of signal 
processsing. 
(3) Image reconstruction using 2-D wavelet. 
According to the Nth level approximation (low-frequency 
coefficient) and all the detail (high-frequency coeficient) after 
threshold qualification, the wavelet of 2D signal is 
reconstructed. 
In these thress steps, the main part is threshold choose and 
threshold quantification. 
Figure 5. Wavelet decompositon flow chart 
Figure 5. Wavelet reconstruction flow chart 
Figure 4 and figure 5 show us the flow chart of wavelet 
decomposition and reconstruction. By use of MatLAB, the 
noisey image is decomposed. However, usual wavelet denosing 
can not remove this stripped noise well (figure 6).Wavelet 
decomposition suggests, stripped noise still exists in the 
horizontal domain and the noise is not removed. However, the 
vertical and cross part of this image does not have much noise. 
So we can decompose this image and denoise the horizontal and 
approximate part, then synthesize the prior image after noise 
reduction. 
In the flow chart, the part of LL and HL will be processed. 
Noise reduction of LL and HL can be low pass filter or other 
methods. Because the detail mainly contains in HH and LH, so 
the detailed information will be reserved much. The results is 
shown in figure 7. 
Figure 6. Image decomposition using Wavelet and usual 
wavelet denoising result (sym4)