Pavement Defect Detection Based on Ground Penetrating Radar and Deep Active Learning

Abstract

Deep learning can assist ground penetrating radar (GPR) to accurately identify the internal defects of airport track structure. However, deep learning often requires many annotated samples. For this purpose, a deep active learning (DAL) method with multiple selection criteria has been proposed for defect detection. In addition, a color data set which found data annotation easier in this case, was created, and the model trained with color images improved performance by about 1% compared with the model trained with the original grayscale images. Subsequently, the selection strategies based on entropy, least confidence, and a combined criterion were proposed as an active learning mechanism. The experimental results indicated that when 443 training images were used, the model using the least confidence-based selection strategy achieved a mean average precision (mAP) of 87.50%, and the model using the combined criteria-based selection strategy achieved a mAP of 87.56%. However, the detection model using entropy-based selection strategy used 393 training images to achieve a mAP of 88.06%, which was superior to the other two selection strategies, and the number of training samples was only 53% of that of the initial model. Overall, DAL could achieve similar model performance to traditional deep learning while reducing annotation costs by 47%. This method has a detection speed of 96 frames per second, which meets the requirements of engineering applications for detecting defects in airport pavements.