A Deep Learning Method for Rough Surface Clutter Reduction in GPR Image

The major clutter in a typical ground penetrating radar (GPR) B-scan image is the ground surface clutter. It can severely obscure or distort the subsurface target response, especially when the surface profile variation is large and the target reflection signal is overlapped with the surface clutter. In this paper, we propose a deep auto-encoder based method for reducing the rough surface clutter in GPR images. In our method, the rough surface clutter reduction is formulated as an anomaly detection problem. The rough surface region in a B-scan image is splitted into small patches which are used as the training data set to train a deep auto-encoder. The trained autoencoder captures the patterns of the rough surface patches. After training, the whole B-scan image is divided into small patches of the same size as the training patches and each of which is fed into the auto-encoder to compute an anomaly score. To reduce the clutter, we design a novel method that adopts a weighted sum approach to aggregate all patches based on their anomaly scores. For performance evaluation, simulation and laboratory tests are conducted, and the clutter-to-rest ratios (CRR) are computed to quantify the clutter reduction effectiveness. The CRR of the simulated B-scan image is reduced from 4.24 dB to -14.06 dB, and the CRRs of two measured B-scans are reduced from 3.71 dB to -20.45 dB, and from 8.62 dB to -8.07 dB, respectively.