Detection of Hyperbolic Reflections on Real GPR Data Using YOLO Model Trained with Synthetic Data

This paper analyzes the application of the YOLO algorithm to automatically recognize hyperbolic reflections on radargrams. A specific shape of hyperbolic reflection on radargrams occurs with objects of circular cross-section that are scanned perpendicular to their axis, such as underground pipes. Requirements for processing a large number of radargrams are increasingly being imposed in practice. Therefore, the development of algorithms for the automated finding of hyperbolic reflections on radargrams has been intensified in recent years. Conventional methods take a lot of time and are more susceptible to noise than algorithms based on deep learning. One of such algorithms is also applied in this paper. The model was trained with 2000 synthetic radargrams generated using gprMax software. The basic idea is reflected in the analysis of the successful detection of hyperbolic reflections on real radargrams using a model trained only with synthetic radargrams. Experimental results show a high level of success.