A Multistream Attention Network for Airport Runway Subsurface Target Segmentation

Abstract

The accurate perception of subsurface objects and defects is vital in airport routine maintenance. For the complex subsurface environment of the airport runway, to obtain the high-performance segmentation of typical subsurface targets, a multistream attention segmentation network is proposed. The network takes the ground penetrating radar (GPR) raw data and the preprocessed B-scan image as multistream input. It carries out sufficient feature fusion on multiple modals, scales, and levels to get robust feature representation. Furthermore, we proposed two attention mechanisms suitable for multistream feature fusion, which can learn more effective features. We validate our method on an actual airport runway dataset. Experimental results show that our method can obtain an F1-measure of 82.08%, 89.12%, and 82.54% for three typical subsurface targets: void, pipe, and steel mesh, respectively.