The Synergy of Siamese Networks and Attention Mechanisms for Enhanced Subsurface Image Matching

Ground Penetrating Radar (GPR) stands as a pivotal non-destructive detection technology extensively employed across diverse domains, including underground defect identification, geological exploration, archaeology, and autonomous driving. Central to its utility is the process of image matching, which remains instrumental in these fields. Conventional approaches to GPR image matching have predominantly relied on manual feature extraction techniques, which are inherently time-consuming and constrained by the subjective experience of the practitioners. Deep learning, in contrast, can extract more discriminative features automatically, coupled with stronger generalization capabilities. This paper embarks on an exploration and enhancement of an existing deep learning model tailored specifically for GPR image matching tasks. Specifically, an optimized Siamese Network architecture augmented with attention mechanisms is proposed and rigorously compared against traditional Siamese Network. The advantages of the proposed network takes the incorporation of attention mechanisms into account which extract soft weighted local information to the network, thus enhancing the end-to-end process. Extensive analysis is conducted using a meticulously collected ground penetrating radar dataset. Empirical results obtained demonstrate that by incorporating the attention mechanism into the proposed system exhibits commendable performance. These findings substantiate the efficacy of integrating deep learning with attention mechanisms in the context of GPR image matching, thereby significantly augmenting matching accuracy. In addition, the findings in this work offers an more robust solution for GPR image matching, which lays the foundation for subsequent research such as localization and navigation for autonomous driving.