Crack Segmentation on UAS-based Imagery using Transfer Learning

Christian Benz, Paul Debus, Huy Khanh Ha, Volker Rodehorst

Abstract: The vast number of images that typically arise from automated structure inspection, e.g. by means of unmanned aircraft systems (UAS), pose a challenge for inspectors. In order to support the human responsible, automated crack detection can help determine grave defects that require close-up inspection. The available crack datasets showed to be unsuited for crack segmentation on UAS-acquired imagery. Thus, a crack dataset was created to reflect the difficulties connected to the UAS-based imagery. The challenges mainly are low resolution and intensity of the cracks and re-occurring planking patterns. A convolutional neural network (CNN) for crack segmentation was derived from TERNAUSNET [1] applying transfer learning. It achieved an average precision (AP) of 75% on the challenging dataset. The code and dataset are available at https://github.com/khanhha/crack_segmentation.