3D InspectionNet: A Deep 3D Convolutional Neural Networks Based Approach for 3D Defect Detection of Concrete Columns

摘要

Deep learning-based defect feature recognition from 2D image datasets, has recently been a very active research areaand deep Convolutional Neural Networks have brought breakthroughs toward object detection and recognition. Due toCNN’s outstanding performance, several recent studies applied it for defect detection in either routine or post-earthquakeinfrastructure inspections and have reported competitive performance and potential toward automating infrastructuresafety assessment. Despite their benefits, the majority of 2D approaches do not leverage or provide 3D depth informationdirectly from the content of the images and as a result do not enable the 3D measurement of severity of these defects. Withthe increased popularity of 3D scanning and reconstruction technologies, there is pressing need for defect recognitionmodels that operate on 3D data. In this paper, a novel framework using Deep 3D Convolutional Neural networks (3DCNNs)termed 3D InspectionNet is introduced to learn 3D defects features from an artificially generated 3D dataset,intended to mimic defects on the surface of concrete columns such as either cracks or spalls. InspectionNet has thecapability of learning the distribution of complex defect features from a large 3D dataset, and distinguishing defectsfeatures. For training 3D InspectionNet, a large simulated 3D defect dataset of 3D CAD models was automaticallyconstructed with labeled defect features. The proposed framework can distinguish defect features from the geometric datasuch as voxels with a high accuracy. The results of this preliminary work demonstrate and emphasize the feasibility andpotentials of this approach for 3D defect detection in automated inspection applications.