Vision-Based Crack Detection of Asphalt Pavement Using Deep Convolutional Neural Network

Asphalt pavement depression, e.g., cracking, rutting and bulges, are the main factors endangering transportation safety and capacity. Detection of these depression is a significant step for pavement management; to date several laser-scanning-based technologies have been implemented for this purpose. However, an automated solution remains a challenging task due to the complicated pavement conditions in real world such as illumination and shadows. In this paper, a vision-based automated detection method for pavement cracks is proposed using deep learning technology, wherein a convolutional neural network (CNN) is trained to learn the features of the cracks from images without any preprocessing. The designed CNN is trained on the image database containing 240 images, based on the open-source TensorFlow framework by Google Brain team, and consequently records with about 96% accuracy. The robustness and adaptability of the trained CNN are tested on 40 images taken from different roads under various crack types, which were not used in the training and validation process. Testing results show that the proposed method has satisfactory performance, and therefore, could be beneficial for providing an alternative solution to automated detection of pavement cracks.