Automated detection and segmentation of concrete air voids using zero-angle light source and deep learning

Abstract The detection and segmentation of air voids in concrete has received significant attention because they are critically important for determining concrete properties and performance. However, previous methods have shown low efficiency and accuracy in void segmentation. Particularly, it remains difficult to detect and segment irregularly shaped voids, especially those that are connected and have indistinct boundary features. This study presents a zero-angle light source to clearly capture features of different types of voids that are otherwise hard to identify using conventional illumination schemes, thus allowing for accurate segmentation of complex air voids by an instance segmentation model using a path aggregation network (PANet). The PANet model significantly outperforms existing semantic segmentation algorithms in detecting air voids, especially small and connected ones. Furthermore, we also show the robustness and generalization ability of this model for air void segmentation and analysis by applying it to different cement-based construction materials.