An automated pothole detection via transfer learning

Abstract

Potholes on the roads can cause many problems in traffic. They can cause malfunctions of vehicles, deterioration of suspension systems, additional repairs, and traffic accidents. It is very important to detect potholes quickly and with low costs for the maintenance and rehabilitation of roads. This shows that there is a need for automatic systems that can detect structural problems that may occur on the roads quickly and accurately. In this study, DenseNet121 architecture, which is a deep learning-based method, is proposed for detecting potholes in roads. With the proposed approach, it is aimed to determine whether there are potholes in the road images in the dataset. In this study, potholes on the road were detected with 99.3% accuracy using the DenseNet121 network. This success is quite high when compared to similar studies in the literature. At the same time, this dataset was run and compared with ResNet50, InceptionV3, VGG19 and InceptionResnetV2 models with the same parameters. Among these models, the highest accuracy was obtained with DenseNet121.