The function reduire_colonnes should also filter in y-direction of the input file, a 2 dimensional median filter would be the ideal choice, as described previously.
It is crucial for the future development of this project to collect as much data as possible. Then, we can try to find algorithms to classify the data. For example, data containing no mines at all can be very valuable training neural networks. This leads us to the list of algorithms to try out in the next part of this project. In [DD95], Daniels and Dittmer discussed many of these approaches.
A finished system that solves all of the requirements stated earlier in the Problems section, is very difficult to realize. Even for humans it is not always easy to interpret GPR created images that were taken in an unknown environment and under new conditions. Also considering the number of different mines, it would be very difficult to find an universal algorithm that handles all of these conditions. A possible solution to this problem would be an adaptive algorithm, which ``learns'' some of its parameters ``on site''. We could imagine to confront our robot first with a terrain that contains no mines at all. Then with a mine on the surface, later mines in different depths would be presented to our robot. During this ``learning'' phase, parameters could be adjusted, to improve the success and false alarm rate at the same time.
The next step also involves to better understand the reflected GPR
signal resulting from a mine. We also need to take pictures, where the
exact position of the mine can be seen. In our images, the distance
measurement wheel of the GPR antenna was sliding on the sand, giving
uneven spaced images. Therefore, we did not know the exact location of
the mine on our images. The future most important step will be to have
a large set ( 
100) of significant images containing mines,
where the exact position is known. Having also a large set of pictures
containing only ground scans is also important, as stated above.