The interesting part of GPR created images are the diffraction
hyperbolae. They can be seen
better by filtering out the horizontal lines on the image. This can be
done by convolving the image with a horizontal high pass filter. The
derivative of a Gaussian represents such a filter. The best results
were achieved with a 10 pixel wide filter, having the shape as in
Figure 9. The filtered image can be seen in
Figure 10. We can immediately spot a great improvement
over the background removal algorithm, as the ground reflection band
is filtered out as well as the horizontal lines. A lot more structure
appears in the image. This filter gave us the best result to improve
the information content. The filter is generated by the Matlab script
hor_remove.m, printed in Appendix A. The Matlab
function conv2 can be used to apply the filter to the image. I
noticed that the filter width is important for the final result. When
using a large window ( 
10 to 20 bit), the results showed
that the contrast of the image was minimized. When the window was too
small for a given picture, the filter introduced noise. Every change
in the image resolution or image size needs also an operator assisted
reevaluation of the ideal filter width.
After the horizontal filtering, it showed to be useful to use a noise reducing filter (because horizontal filtering is a derivative action, which introduces noise). A 2 dimensional median filter yielded the best result. We used the following Matlab commands for our tests
I_hr = conv2( I, filtre_hor, 'valid' ); I_hrf = medfilt2( I_hr, [5 3] ); I_hri = conv2( I_hr, filtre_ver, 'valid' );
I contains the reduced image (200 by 512 pixels). I_hr
designates the horizontally filtered image. I_hrf is obtained by
filtering I_hr with a 5 by 3 median filter
. Filtre_ver
contains a 1 by 5 pixel vertical filter, created by a Gaussian
function, which is used to average over 5 pixels in order to reduce
the noise in the vertical direction. Our
experiences showed that the median filter was
superior. Figure 10 shows an image where the Median
filter has been applied to.
Figure: 10 bit wide horizontal filter
Figure 10: sand3 with Horizontal
Line Removal and Median filtering