Automation asphalt pavements distress detection using deep learning-based RetinaNET

Keywords

Automated crack detection, asphalt crack detection, deep learning, RetinaNet

Abstract

In this paper, the author proposed a supervised machine learning network to identify and classify different crack types in asphalt-surface pavements. A laser camera captured surface images from pavement surface then classified them into 3 classes following the manual of pavement distress identification by the Federal Highways Administration (FHWA). These classes are three different crack types: alligator (fatigue), longitudinal, and transverse cracks. The training database was collected from 1,000 images with the original size of 3,704x10,000 pixels. These images then were divided into 20,000 smaller images of 1,852x1,000 pixels size. The images data are labeled based on the nine crack types and trained using a deep learning algorithm called RetinaNet. The trained model is verified using 2,400m of pavement surface images obtained from Seoul city urban road. The results have shown that the trained network model has an accuracy of around 85% for crack detection and classification.

Downloads

Download data is not yet available.