U-Net With Dilated Convolution for Improved Clutter Removal in GPR Images

The clutters are inherent in ground penetrating radar (GPR) investigations which severely affects the performance of subsurface target detection. Therefore, removal of clutter is an essential preprocessing step in GPR survey. Several approaches are existing in literature for this purpose. However, it is quite challenging to remove the clutters for a rough ground surface using conventional methods. In this paper, a dilated convolution based U-Net architecture is proposed to improve the clutter removal performance of GPR B-scan images. The effectiveness of proposed approach over standard U-Net is validated on synthetic as well as laboratory measured data through qualitative and quantitative evaluation.