Contribution

SIC is responsible for individual sensor data processing and data fusion. Some geo-coding has also been performed.

Methodology

We concentrated our effort on the development of tools to accelerate the work of photo interpreters by finding region of interest. For that, SIC's graphical interface has been enhanced so that a user can easily navigate through the image and take the final decision.

During phase I, we decided not to use any contextual information, mainly because this information may not be relevant in a real case. We therefore selected mine indicators and built some algorithms to detect them. Since where mines are laid some vegetation anomalies can be seen, we focused on the detection of "easy-to-find" characteristics, typically using the spectral response of a mine laid on the ground.

At the end of phase I, photo interpreters gained an experience that we decided to use. We are now working on new algorithms based on methods used by photo interpreters during phase I. This will help reducing the false alarm rate.

Images used

A near infrared scanned image from MK2000 camera provided by Aerodata and covering minefield C.

A colour scanned image from Leica camera provided by Eurosense and covering the neighbourhood of minefield C.



A colour infrared scanned image from Leica camera provided by Eurosense and covering the immediate neighbourhood of minefield C.



After the Belgium tests some additional images were available:

• Radar (X and P band) : around 30 images provided by Aerosensing
Not geo-coded and difficult to geo-code by SIC
low resolution for X-band
• Thermal infrared: about 25 images provided by Recon Opticals
Not geo-coded and difficult to geo-code by SIC
• Digital colour: 7 images at the time from Vos80 camera provided by EOS
not geo-coded and difficult to geo-code by SIC
huge distortion
The quality of the images acquired before the Belgium test is far better. Therefore they were used for automatic analysis.

Graphical user interface

The images were visually inspected to find features characteristic of mines. Because the available images are huge, not registered and coming from different sensors at different resolutions, it was necessary to enhance our graphical user interface to allow an efficient navigation through the available image database. Using new tools, most of the images were geo-referenced, and put together into a virtual image called MIF (multiple image format). The position of the known mines was encoded together with their attributes. All this allows for an easy navigation trough the known mines. For a recorded mine of interest the user may request the available attributes, such as type or depth, as well as the available image channels. It is then possible to go from an available channel to another or to display all channels simultaneously in a common projection.

Building a mine (indicator) model

To develop any tool for semi-automatic mine detection, a model of a mine, or a mine indicator, is needed since all such tools encode a model either explicitly or implicitly. The expertise of photo interpreters is very useful to build these models. The model presented here was designed by SIC without any input from photo interpreters.
Some vegetation anomalies (missing vegetation for instance) can be seen where mines are laid. Unfortunately, a huge amount of such regions exists and to reach a reasonable false alarm rate, complementary mine characteristics (specific patterns, signature in different spectral bands, etc.) have to be introduced in the model. Therefore, in a first approach the typical spectral response (Visible and near infra-red) of an AT mines lying on the ground was used.

Algorithms for mine detection

Two algorithms have been implemented for mine detection.

The first algorithm is based on a maximum detection followed by a region growing in the blue band. The growing algorithm uses a rough model of the mine to compute a local contrast that is then used to derive the growing stop criteria.

The second algorithm, written by Vinciane Lacroix, is based on a threshold in the blue band to find regions of interests. Afterwards, edges are computed in those regions and grouped. Finally, the shape of the region is considered and the candidate is kept if it is sufficiently circular and has an appropriate radius.

Detection of partially occluded circles

As explained above, the shape of the candidate regions may be used as discriminant feature. Mines lying on the ground present circular shapes. However, the mine is often partially occluded and we developed an algorithm that is able to recognize circular shapes even if they are partially occluded. A measure of the circularity is returned together with an estimation of the radius. The radius provides a better estimation of the size of the object than the visible area (partial occlusion) and is thus more discriminant.

 The algorithm estimates a circle that goes through most of the pixels of a given list. This is useful when it is known that the pixels should lie along a part of a circle but some of them can be wrong and very far from the circle.

Several triplets of pixels are selected in the list and the circles defined by these three pixels are computed. For each circle, the distances of all the pixels of the lists to this circle and the percentiles of these distances are computed. The circle selected is the one giving the smallest percentiles of errors. The number of triplets is computed so that the probability to have at least one triplet without erroneous pixel, and thus a correct circle, is higher than 0.99.

Close-up of an anti-tank mine in colour infrared. Only a part of the mine is visible. The blue colour is due to the fact that the colour green is seen blue in colour infrared.

Contour of the mine detected. Because of the occluded part, the shape is not circular.

Circle estimated from the previous contour. It is a good approximation of the position and size of the mine. If an absolute co-ordinate system is available, it is possible to have an approximation of the real size of the mine and then an indication of which type of mine we are dealing with.

The following images present the result of the automatic mine detection on the full visible and colour infrared images.

On the visible image, a V-shaped minefield has been detected. Note that no mines of minefield C has been found. The anti-tank mine 281 (anti-tank mine lying on the ground and clearly visible) was first detected but rejected by the attribute based filter. Using a better filter criterion, it could be possible to keep that mine without increasing the false alarm rate but little effort was spent for this fine-tuning because we believe that an algorithm that learns from example should be used in practice. This could not be tested until now because the database contains too few examples of known and detected mines. In this context a learning scheme would probably lead to poor results (over-training = learning by heart rules that are only valid for the training set and that may not be extrapolated to new data).
 
 

On the colour infrared image representing minefield C at a higher resolution but with only a small neighbouring region (the V-shaped minefield is not in the imaged region), three anti-tank mines (on and below the surface) of minefield C have been detected. Note that for the buried anti-tank mines, it is a stick lying near the mine that is detected. Even if the stick may be considered as a valid mine indicator, they were found by chance because the algorithm was not developed to find such objects. We concentrated our efforts on the detection of indicators that could be helpful in real cases. If it appears that such sticks are good indicators, a better detector will be developed. Note that the size of the visible image is about 400 MB and that this image covers a region of about 500 square metres. The full image was processed in about 30 minutes on a Pentium Pro (200 MHz, 64 MB RAM) running under Linux. The position of known mines has been superimposed on both images (dots are known mines and crosses are mines proposed by the algorithm).

The tool was developed to find anti-tank mines lying on the ground. Most of those mines were detected and the false alarm rate is reasonable. Even two buried mines were found because a stick left in the vicinity of the mines was detected. We believe that the false alarm rate could be reduced in the near future by some simple improvements. As an example many false alarms in the colour infrared are found in the trees. A tree detection algorithm could reduce the false alarm rate significantly.

 Of course, if image processing is only able to detect anti-tank mines lying on the ground, its usefulness in real situation would be quite limited. However, if photo interpreters are able to detect more minefields using other mine indicators, image processing could be helpful to detect those indicators leading to a substantial acceleration of the data analysis.

Although limited, the results of the first evaluation presented here have shown that significant process acceleration can be reached by means of image processing. By removing the false alarms found in the trees of the colour infrared image, the found alarms could be grouped in about 10 regions of interest. If the photo interpreter takes 30 seconds to look at each region, 5 minutes would be needed to analyse the scene. Whereas a full visual inspection was carried out at full resolution, the photo interpreters would typically spend at least one minute for a region of 1,000 on 1,000 pixels. 150 such regions have to be analysed for a full coverage of the colour infrared image leading to about 2 hours of interpretation. A process acceleration of about 60 may thus be expected. Such a speedup would be very useful in an operational context where the amount of data to be analysed would be tremendous. Without image processing and with the same assumptions as above, 650 man-hours would be needed for a complete visual inspection of a scene of 10 square kilometres.

For comments or questions contact: Pascal Druyts or Yann Yvinec.