Pop-outs Segmentation for Concrete Prognosis Indices using UAV Monitoring and Dense Dilated Convolutions

Abstract

Efficient inspection and accurate prognosis are important issues for aging concrete structures. In particular, volcanic aggregates that contain moisture can freeze and expand, resulting in "pop-outs". This pitting corrosion is sparsely distributed; and hence, frequently occur in small regions on the surface of concrete structures. For this reason, it is difficult to accurately count pop-outs and measure their size. If we can automate the prediction of the pop-out region, then repair targets can be set for concrete prognosis. In this paper, prototype concrete segmentation tool for application to sparselydistributed pop-outs was developed to extract damage features from drone images. We propose dilated convolutional networks with several four- and eight-branch atrous sampling rates. Specifically, we applied the methods to 40 drone-monitored images with size of 6,000×4,000, partitioned into 16,000 unit images. We computed pop-out prognosis indices to support repair actions that make sense for dam managers.