GPR-based hand-held landmine detection methods using B2D-DCT-PCA and conditional-GAN

Abstract

In this dissertation, we propose two landmine detection methods applicable to ground penetrating radar (GPR)-based hand-held landmine detectors. The first method can be used to detect landmines quickly and accurately in situations wherein the amount of training data available is small. Conventional hand-held landmine detectors have the advantage of being able to detect landmine signals quickly

however, they also have the disadvantage of being sensitive to non-target signals such as land unevenness, sensor movement, and noise signals. To achieve robustness against background and noise signals, the landmine signal must be detected by comparing it with a precisely generated background model

however, the amount of computation required in this case significantly exceeds that required in conventional methods. Therefore, we propose a bilateral two-dimensional discrete cosine transform-principal component analysis (B2D-DCT-PCA) that achieves robust detection against background and noise signals by using background models. Furthermore, our proposed method alleviates computational burden by reducing the dimension of the features. In addition, to improve the detection performance of the proposed method, the landmine-background reconstruction error (LBRE) is defined to find a suitable pre-processing method. Frequency shifting is then chosen as the pre-processing method because of its superiority in terms of increasing the LBRE. The detection performance is further improved by adding a post-processing operation, namely, multiplying the weight values according to the depth to compensate for weak signals deeper in the ground. Experimental results showed that the proposed method outperforms conventional detection methods, and the processing speed is higher than that of the one-dimensional PCA method as well as of the proposed method without the two-dimensional DCT. The second method uses conditional generative adversarial networks (cGANs) to estimate the location and size of buried landmines more accurately. Assuming a landmine detection situation in practice, the networks cannot be trained using ground truth data because information regarding which landmines or explosives are buried is unavailable. Therefore, we propose a method that has shown excellent detection performance by using cGANs based on a distance map derived from B2D-DCT-PCA when no ground truth data is available. Because many false alarms are included in the distance map, to improve detection performance, the networks must be configured such that their outputs have fewer false alarms than those of the distance map. Therefore, the usage of a combination of both false alarm loss and intersection over union loss as the loss function of the networks is proposed, in addition to a method that simultaneously uses raw GPR data and pre-processed GPR data to train networks. In other words, the goal of the proposed method is to generate data close to ground truth data by reducing false alarms as well as by generating images similar to distance maps. As a result of performing experiments according to the application of each proposed method, it was confirmed that using both the proposed methods yielded excellent detection performance. In addition, it was confirmed that detection using the proposed cGANs was superior to using distance maps derived from B2D-DCT-PCA.