Deep Learning Neural Networks for sUAS-Assisted Structural Inspections: Feasibility and Application

Abstract

This paper investigates the feasibility of using a Deep Learning Convolutional Neural Network (DLCNN) in inspection of concrete decks and buildings using small Unmanned Aerial Systems (sUAS). The training dataset consists of images of lab-made bridge decks taken with a point-and-shoot high resolution camera. The network is trained on this dataset in two modes: fully trained (94.7% validation accuracy) and transfer learning (97.1% validation accuracy). The testing datasets consist of 1620 sub-images from bridge decks with the same cracks, 2340 sub-images from bridge decks with similar cracks, and 3600 sub-images from a building with different cracks, all taken by sUAS. The sUAS used in the first dataset has a low-resolution camera whereas the sUAS used in the second and third datasets has a camera comparable to the point-and-shoot camera. In this study it has been shown that it is feasible to apply DLCNNs in autonomous civil structural inspections with comparable results to human inspectors when using off-the-shelf sUAS and training datasets collected with point-and-shoot handheld cameras.