Inspections of Crack Surface on Irrigation Structure Based on Structure from Motion Technique and Using Deep Learning by Fully Convolutional Network

Abstract

Currently, the Royal Irrigation Department perform visual inspection on structures in their regular maintenance routine. All elements are required to be inspected again Condition Index (CI) to assess the current state of dams. However, the visual inspection method cannot be performed thoroughly, as there are many areas that are difficult or inaccessible for inspectors, often some areas can be harmful to surveyors. In addition, the visual inspection method posts significant amount of errors that are caused by the inspectors themselves.To ensure the structural integrity of a dam, in this paper, the novel method of inspection is proposed using photographic technology through high-resolution cameras. An unmanned aerial vehicle are used for data collection, and then a 3D model of a dam is created through photogrammetry. Then the Fully Convolutional Network (FCN) technique is applied to detect cracks on the dams. The results show that the proposed system can detect cracks at pixel-level with accuracy over 90%.