AF-NET: Urban Road Vehicle Ground Penetrating Radar Image Denoising via Attention U-Net

In urban environment, vehicle-mounted GPR images of urban roads may be disturbed by various noises, which will affect the correct recognition of underground structures. Therefore, de-noising radar images can improve data quality, ensure accurate detection of underground structures, and optimize urban planning and infrastructure management. However, there are few data sets of vehiclemounted GPR images on urban roads, and there is a lack of corresponding models for noise recognition. To solve these problems, an innovative radar image denoising framework, namely AF-NET, is proposed. This framework is based on convolutional neural network, uses Attention U-Net as the core framework, and adds LayerNorm, SCA and Simple Gate to improve computing efficiency. Experiments show that the collected radar image data set exceeds the previous SOTA0.25 dB. In the experiment, we successfully de-noise the radar image by designing the training image codec and using the features extracted by AF-NET. The PSNR of this method on the acquired radar image dataset is 43.34 dB, which verifies the effectiveness of AF-NET in the radar image denoising task.