Enhancing anomaly detection in ground-penetrating radar images through reconstruction loss and high-variability

The aim of this study is to enhance the anomaly detection capabilities of ground-penetrating radar (GPR) images by adopting a novel loss function composed of cross-entropy and reconstruction loss. The reconstruction loss measures the error between the inputs and reconstructed inputs via a network that reconstructs inputs using the features extracted by the classification networks. Additionally, GPR images with high variability are generated by combining raw GPR images with background noise collected in the survey fields and white noise (Gaussian noise). The experimental results show that activating reconstruction loss results in an increase in training time (approximately 1.79 times) but significantly improves anomaly detection performance (accuracy peaking at 92.5 %) with the utmost optimal weighted factor being 0.1 using various state-of-the-art classification networks. Furthermore, simulating highly variable GPR images using background and white noise significantly improves the detection accuracy. Background noise introduces noisy details into the GPR images, whereas white noise functions as a high-pass filter depending on the coefficient of variation. This study suggests that the proposed loss function and image manipulation technique can effectively enhance anomaly detection performance.