Abstract
Maintaining safe and reliable roadway infrastructure is a critical challenge that demands constant monitoring and analysis of surface level pavement distresses. Typically, this maintenance involves identifying and quantifying various forms of road damage, such as cracks and potholes, which are indicative of the pavement's overall health and safety. Recently, deep learning (DL) based automated methods have been recognized as the state-of-art for pavement distress analysis. These methods streamline the maintenance process through a two-step procedure: initially localizing areas of distress on the pavement (i.e., through object detection models) and subsequently performing pixel-level segmentation to quantify the severity of the damage (i.e., through an image segmentation model). However, the effectiveness of DL models is significantly hampered by feature-level distribution shift, a common problem where there is significant difference between training data and real-world data in terms of features like brightness, contrast, texture among other statistical features. This issue affects DL model's generalization ability, limiting its accuracy on new or unseen data. This paper introduces an innovative and cost-effective approach to enhance model generalization in the context of pavement distress segmentation. The proposed solution centers around an unsupervised generative data augmentation strategy that transforms features of new or unseen data to align closely with the training dataset before performing distress segmentation. The framework's effectiveness in improving pavement distress segmentation ability, is demonstrated through comparative analysis against traditional methods under varying distribution shift scenarios. Results indicate a significant improvement in segmentation accuracy, highlighting the potential of generative data augmentation strategy to address distribution shift challenges. This paves the way for future advancements in pavement distress analysis and model generalization.