Machine learning for improving metal detection

Locating and identifying hidden conducting objects has a range of important metal detection applications in the aerospace and defence sector including identifying landmines and the early detection of concealed terrorist threats. There are further applications in the built environment (eg identifying hidden pipelines and cables) as well as in archaeology (e.g searching for buried treasure). While metal detectors do detect metallic objects, they have difficulties in distinguishing between their shape and, thus, have issues in identifying exactly what is hidden. For example, distinguishing between coins and keys accidentally left in a pocket during a security search and a person carrying a security threat. Or a treasure hunter wishing to distinguish between old coins and some old scrap metal that is buried underground. Furthermore, there are further challenges when multiple objects are present.

The aim of the PhD will be to improve the identification and location of hidden conducting objects from metal detection measurements using machine learning techniques. The project is aligned with the EPSRC research topic classifications of Numerical Analysis, Artificial intelligence technologies and Sensors and instrumentation and the EPSRC industrial sector classifications of Aerospace, Defence and Marine and the Built Environment. The specific objectives of the work are

1. The student will contribute to the development of a new methodology and a new Python implementation of a computer simulation tool, based on the numerical analysis techniques of reduced order models and finite element methods, that allows conducting objects to be described using a small number of parameters.
2. The student will compute a new dictionary of object characterisations by creating computer models of different geometries, applying the tool in objective 1 and studying the effects of numerical discretisation on the characterisations of relevant objects for security screening.
3. The student will investigate and tailor supervised machine learning techniques to the classification of objects using the object descriptions developed in objective 1 and the dictionary developed in objective 2. This will involve first comparing the performance of different machine learning technique in existing libraries (eg Skikit-learn) for this application by writing new Python code to interface with his dictionary and making new studies of performance using cross validation and other metrics. Then changes to the techniques will be proposed so that they can be improved for his application in order to decide the best classification approach.
4. The student will apply the classifier developed in objective 3 to the identification and location of objects using field measurements from metal detection sensors.

The novel aspects of this work include the development of a new computer simulation tool as well as the underlying numerical analysis methodology, the development of new library of object characterisations, a new understanding of the effects of numerical discretisation on these characterisations and the tailoring of machine learning techniques to this application.