A comparison of neural networks and sub-space detectors for the discrimination of low-metal content landmines

Low-metal content landmines can be particularly difficult to detect and classify with electromagnetic induction (EMI) systems. Their responses are often less than that of indigenous clutter and the small amounts of asymmetrically distributed metal results in significant changes in the signature of the mine as the sensor to target orientation varies. A number of algorithms have been previously developed in order to aid in target classification and reduce the false-alarm rate. In our work, multiple data sets were collected for each of five targets, of varying metal content, at several sensor to target heights and horizontal displacements using a prototype frequency-domain EMI sensor, the Geophex GEM-3. The data was then evaluated using one of three classification algorithms including a neural network, a matched filter, and a normalized matched filter. Here, a One Class One Network (OCON) architecture in which only one neural network makes a decision was selected for use. We will discuss the training and testing process for this algorithm. We will also show that the neural network performed much better than the matched filter but slightly worse than the normalized matched filter. In addition, the results demonstrate the necessity of training the algorithms with spatially collected data when precise sensor positioning is not possible.