DETECTION OF CEILING SAGGING BASED ON DEEP LEARNING OF IMAGES AND POINT CLOUDS

Abstract

Deep learning is applied to images and point clouds data acquired from specimen simulating ceiling sagging. For deep learning on images, AlexNet’s classification accuracy from transfer learning is 93.4%. Principal component analysis of the point clouds is used to determine the curvature of the point clouds. By setting curvature threshold for classification, correct rate is 96.5%. In deep learning for point clouds, classification accuracy by PointNet is 96.7%.