T&A, Airvision* and the Universities of Eindhoven and Tilburg are conducting joint research to determine whether the use of Artificial Intelligence (AI)/Neural networks can contribute to a more efficient detection process for explosives from the Second World War.
The usual way to analyse magnetometer data involves the following steps:
1. Search the measurement data for locations with high amplitude values.
2. Using inversion of the measurement data based on theoretical deterministic formulas, determine whether these amplitude values can be caused by an explosive at which the location is suspicious.
3. If this is the case, the object is added to a list of suspicious objects that need to be accessed.
In daily practice, it appears that the drawn up list of suspicious objects contains a large number of false positives. These are therefore suspicious objects that do not appear to be explosives when visually identifying the suspicious objects. This therefore results in additional costs.
We are now investigating whether we can optimize steps 2 and 3 with a neural network procedure. The results of approach work are used, which are then compared with the drawn up list of suspicious objects.
The first tests show encouraging results. In one of the projects it turned out to be possible to reduce the number of false positives towards 50%. This means that almost 1 in 2 suspicious objects also turns out to be an explosive when visually identified. However, the downside of AI was also that 6% of the explosives would have been missed when using it. Explosives that were found during the regular method of investigation.
However, the use of AI is in full development and further improvement can be expected by using more reference data. The investigation into this will continue.
*Airvision is a start-up that developed a no-code platform for organizations to create tailor made computer vision AI models without needing technological expertise. Organisations can upload drone imagery, annotate images and train world class models. Everything in one platform, fully optimized for drone imagery.
