A GPR Resolution Enhancement Method Based on Weakly Supervised Learning

Abstract

Ground penetrating radar (GPR) is mainly used to detect underground structures in archaeology, military and other fields. In practice, the detection depth and resolution of GPR data are often difficult to balance. Therefore, it is vital to develop a method that can obtain both deeper distance and higher resolution. In this paper, a weakly supervised method for improving the resolution of GPR data based on the Cycle-Consistent Adversarial Network (Cycle-GAN) is proposed, which improves the quality of GPR data by learning the mapping relation between low-resolution and unpaired high-resolution data. The actual data is used to verify the validity and feasibility of the proposed method. Experimental results have shown that our method is able to recover detailed high-frequency components and the resolution is effectively improved.