Author(s) :

Nathan Lin Xiao, Wei Cheng.

Volume/Issue :

Abstract :

Roadway distress detection is essential for ensuring a safe and comfortable driving environment. However, given the irregular shape, small area size, and occasionally very large number, of the road distress objects, it is often laborious to label the distress instances during the training process under the fully supervised algorithm. To address this issue, the study strives to apply semi-supervised learning for distress detection that claims to reduce the cost associated with the labeling process, while maintaining or even improving the learning accuracy in some situations. The research features three distinct backbones of Mask R-CNN models, Unmanned Aerial System imagery of two resolutions, three levels of pseudo-labeled data, eleven threshold values and two types of assessment (that is, in-resolution and out-of-resolution). The results demonstrate that semi-supervised Mask R-CNN models are effective in detecting road distress. Nonetheless, the sensitive analysis is recommended in the future research to identify the optimal pseudo ratio that could generate the highest prediction accuracy.