Structural Damage Semantic Segmentation Using Dual-Network Fusion

Abstract

Road cracks seriously affect road life and driving safety, and road crack segmentation has been a key task in many research fields and industries. Scenes of existing road crack datasets are relatively homogeneous, most datasets only have local crack images taken from a single perspective, and the images lack complex interference pattern other than the cracks and the road itself. However, real-world images of roads are much more complex, and roads may have other dark patterns that interfere with crack segmentation. Perspective and lighting conditions may also be different between every image in the dataset, which makes many road crack segmentation methods extremely low accuracy and severely affects the generalization ability of the semantic segmentation model, while many applications require higher precision rather than recall. In this paper, two methods are proposed to improve the accuracy: augmenting complex patterns in road images and the fusion of classification networks. The methods greatly improve the performance of road crack segmentation.