Ceiling damage detection and safety assessment in large public buildings using semantic segmentation

Abstract

To solve the limitations of traditional on-site inspections by professionals, an automatic method using the semantic segmentation network Deeplabv3+ with transfer learning (TL) is proposed for rapid detection and safety assessment of damaged ceilings in large public buildings after an earthquake. A total of 1389 ceiling images of various scenes are used as the dataset after being manually labeled and divided into four ceiling types: intact, fall-off, suspension, and crack. The model performed excellently in damage detection and location with a pixel accuracy (PA) of 97.23% and a mean intersection over union (mIoU) of 84.18%. Considering the dangerousness of different damage types, the interaction between different damage types, and the ceiling fall-off rate calculated by semantic segmentation results, the safety of the ceiling region is assessed with an accuracy of 98.56%. Some influences are discussed to enhance the detection and assessment ability of the model.