Classification of buried metal objects using wideband frequency-domain electromagnetic induction responses: a comparison of optimal and sub-optimal processors

Classification of buried metal objects can be important for metal landmine and unexploded ordnance (UXO) detection. Previously, the authors have investigated classification of landmine-like metal objects using wideband frequency-domain electromagnetic induction (EMI). The classification performance of a Bayesian detector that incorporates modeled wideband EMI signatures as well as sensor/target orientation uncertainties was compared to that of a processor reminiscent of a matched filter bank. In this suboptimal, but fairly standard approach, each of the filters is essentially 'matched' to the response obtained at the mean position of one of the targets. Both simulated and measured data were considered. The measurements were taken using a prototype wideband frequency-domain EMI sensor, the GEM-3. Dramatic performance improvement was obtained by implementing the optimal classifier. However, the optimal processor is not usable for real-time applications, as it is computationally expensive to perform the integration over the uncertainties associated with the target/sensor orientation. The authors examine the performance of alternative processors whose computational complexity is similar to that of the matched-filter bank, but which take advantage of the physical nature of the wideband EMI response. The wideband frequency-domain response of targets with substantial metal content changes substantially in overall level as the target/sensor orientation changes, but the overall structure of the response as a function of frequency changes only slightly. To exploit this structure in a suboptimal processor, the output of the sensor is normalized prior to processing. Using the normalized and unnormalized data, the optimal classifier performs better than the matched filter classifier. However, the discrepancy between the two is much smaller for the normalized data, indicating that normalization can mitigate the uncertainty associated with the target/sensor orientation. Interestingly, the optimal classifier operating on the normalized data outperforms the optimal classifier operating on the unnormalized data.