| dc.description.abstract | Image-based systems are becoming popular to collect pavement condition data for pavement management activities. Pavement engineers define various distress categories based on pavement types. However, software solutions today have limitations in correctly recognizing pavement types from the collected images in an automated way. This paper presents a convolutional neural network (CNN)-based PvmtTPNet to automatically recognize pavement types at acceptable levels of consistency, accuracy, and high-speed. Pavement images on asphalt concrete pavements, jointed plain concrete pavements, and continuously reinforced concrete pavements in varying conditions were collected via the PaveVision3D system in 2018. A total number of 21,000 two-dimensional (2D) images were prepared, while 80% and 20% of them were randomly selected for training and testing. The CNN network included six layers with 992,979 tuned hyperparameters and achieved 99.85% and 98.37% prediction accuracies for training and testing in pavement type recognition. Images obtained from another two data collections in 2019 were used to validate the PvmtTPNet, and 91.27% and 96.66% prediction accuracies were reached, individually. In addition, the PvmtTPNet shows the highest precision, recall, and F1-score for asphalt concrete (AC) images, which is followed by jointed plain concrete pavement (JPCP) and continuously reinforced concrete pavement (CRCP) images. The developed methodology can provide substantial assistance toward a fully automated pavement condition data analysis for image-based systems, even though a near 100% accuracy is the final objective of the continuing research. | |
