A Deep Learning Assisted Mean-Enabled Laplacian U-Net for Clutter Removal in Complex GPR Images With Reduced Computational Load

Ground Penetrating Radar (GPR) is one of the important geophysical tool for non-destructive subsurface investigations. However, the GPR data is often contaminated by clutter, which can significantly hinder the interpretation of subsurface features. Hence, its elimination is crucial for further GPR data processing. Numerous approaches exist in the literature to address this problem. Traditional methods often exhibit certain drawbacks and suffers from the generalizability issue across diverse scenarios. Deep learning techniques have emerged as a solution to overcome these drawbacks. However, many deep learning methods demand considerable computational resources. In this paper, we have proposed a deep learning assisted lower-complexity network known as Mean-enabled Laplacian U-Net for clutter removal in GPR images. The base U-Net model is popular for semantic segmentation tasks due to its ability to capture both local and global features. In order to further improve its discriminating ability, Laplacian filters are applied on the mean of the encoder’s features along the skip connection path. This integration aids in extracting the essential information for clutter identification and its removal with a reduced computational burden. The proposed approach’s effectiveness has been tested on both synthetic and measurement data through qualitative and quantitative evaluation. It is observed that, the proposed method demonstrates superior performance compared to existing state-of-the-art approaches, achieving the highest average PSNR of 37.5825 and with a FLOP count of 5.8092 billion.