GPR-based landmine detection using histogram dissimilarity and conditional generative adversarial nets

Abstract

In this dissertation, we propose two landmine detection methods using ground penetrating radar(GPR). The first method is to strengthen the landmine signal effectively, even if the landmine signal is weak and there are fluctuations and misalignments in the background signal. The background signal is modeled by using the cumulative intensity distribution(CID) to solve the fluctuation and misalignment problem of the background signal. Before modeling, the frequency shifting is applied to remove the horizontal striped pattern noise which strongly affects the intensity distribution. For quantitative evaluation, we propose a landmine-background distance based on the linear discriminant analysis(LDA). As a result, we can see that the proposed method effectively strengthens the signals in the landmine area. The second method uses generative adversarial nets(GANs), a kind of deep neural network, to convert the mine signal into a ground truth image. To this end, we first propose gray level ground truth, which was inspired by soft targets used in knowledge distillation. Gray level ground truth is to model the target value with a gray level signal, without modeling the ground truth for a binary signal with a target value of 1 and a background value of 0. The target value was modeled by analyzing the tendency of the landmine signal, and the system was observed to converge stably when the gray level ground truth was used. In addition, we apply two ideas to improve the performance of the system. The first idea is the overlap score loss based on conventional performance measurement method of landmine detection systems. This loss differs from the l1 loss in that it takes into account the size of the overlapping area. The second idea is joint training which uses signals from the ambient sensors. This can increase the detection rate of landmines across multiple sensors and reduce false alarms caused by transient noise. Experimental results show that network performance improves when using the proposed ideas. Also, when the results of the two landmine detection methods are merged, the system exhibits the best performance below a certain false alarm rate.