Missing data reconstruction based on an untrained deep decoder network for ground-penetrating radar

Ground-penetrating radar (GPR) data quality significantly affects the precise analysis of physical properties and geometric structures of underground targets. Due to detection environments, instrument performance, poor antenna coupling caused by surface undulations, and a mismatch between sampling interval and sampling speed during the data collection, the obtained GPR data inevitably suffers from random or regular missing signals or traces. Recently, GPR missing data reconstruction methods based on deep learning (DL) have received widespread attention. However, most existing methods require abundant training data samples for model training, which are difficult to obtain in the GPR field. To solve this problem, we propose a data reconstruction method based on an untrained deep decoder network, which requires only one image and does not need training as in conventional DL-based methods. The reconstruction performance is evaluated using laboratory data and field data. Visualization and quantitative results demonstrate that the proposed method can successfully recover the B-scan for column-wise and pixel-wise missing data cases, even at a high missing rate. Moreover, the proposed method outperforms the deep image prior and conventional matrix completion methods.