A denoising method based on cyclegan with attention mechanisms for improving the hidden distress features of pavement

Abstract

Ground Penetrating Radar (GPR) is one of the most used devices for road structural damages detection. However, due to the different roadbed conditions and various disturbances in the nearby environment during detection, there are great difficulties in interpreting detection images, which also hinders automatic detection based on deep learning. In this work, we design a GPR image denoising method based on Cyclegan. We select the most suitable generator and add different attention mechanisms. After denoising the natural GPR road detection image, using the Yolo (You Only Look Once) to test the accuracy of the original image and the denoised image after adding different attention mechanisms. The detection accuracy is improved by 30%. The results of the detection network and the evaluation of the denoised images by GPR image interpreters indicate that the method has the following advantages: lower requirements for training data sets, a wide range of data sources, low cost, good denoising effect, and automatic detection of GPR images. It is of great help to the automatic detection of GPR images.