Faster region convolutional neural network for automated pavement distress detection

Abstract

Pavement images have been utilised to detect distresses. However, existing methods for detecting pavement distresses are not acceptable owing to the various real-world conditions. To complete the task, a novel detection method based on faster region convolutional neural network (Faster R-CNN) was utilised to recognise and locate pavement distresses including crack, pothole, oil bleeding and dot surface autonomously. Twenty Faster R-CNNs were trained and tested by 6498 pavement images. Then the performance of training was analysed to select the optimal Faster R-CNN. At last, a test and a comparative study were presented to verify the stability and superiority of the optimal one. In the testing, average results of the accuracy rates, recall rates and location errors in the optimal one were 90.4%, 89.1% and 6.521 pixels, which were close to average results of training and validation. It indicated the optimal Faster R-CNN had good performance to detect distresses in different pavements. Compared with the CNN and K-value method, the optimal Faster R-CNN located pavement distresses with bounding boxes more precisely.

Keywords:

• Pavement distress detection
• faster region convolutional neural network
• region proposal networks (RPNs)
• automated detection
• pattern recognition
• pavement image

Additional information

Funding

This work was supported by industry key technology project of the Ministry of Transportation of the People’s Republic of China: [Grant Number 2018-MS1-025]; Technology project of Xinjiang Transportation Department: [Grant Number 2018-6].