The original images we took at EMPA, were 2100 by 512 pixels with a 16 bit resolution. This size is too big to be efficiently treated by Matlab. Further, reducing the size of the image can be used to average the data and reduce the noise. We wanted to reduce the size only in the horizontal direction, preserving the depth information. It seemed reasonable to work with 210 x 512 pixel images, which results from a reduction of the horizontal pixels by the factor 10.
The first method we used took the arithmetic mean of the 10 pixels. This filter is linear and is responsive to stray pixels with high intensities. We therefore used a non linear filter instead, called median filtering. This method sorts the 10 input pixels and takes the mean between the 5th and 6th pixel (the middle pixel in the odd case). The result can be seen in Figure 7. The median filter performs very well in the case of binary noise or stray pixels. The Matlab code for the reduire_colonnes function can be found in the Appendix A.
Another pretreatment used to increase the contrast of the image was background removal. The strong ground reflection results in a high intensity wave on top of the picture. The contrast of this wave is high, such that the other activations on the image can not be seen. Subtracting the mean of a line of input pixels from each pixel on the same row, gave us some sort of background removal. The result can be taken from Figure 8. We can spot immediately that the ground reflection is mostly eliminated. The color range is better used and the first interesting structures appear in the middle of the image. This algorithm can be further improved by subdividing the image into smaller bands (about 50 pixels), and applying the same algorithm. The idea behind is that the background also changes over the image and that this subsampling method is more responsive to the ``local'' background. The Matlab program implementing this function is included in Appendix A, called background_removal.m.
Figure 8: Sand3 with Background
Removal