Real-Time Ground-Level Building Damage Detection Based on Lightweight and Accurate YOLOv5 Using Terrestrial Images

Abstract

Real-time building damage detection effectively improves the timeliness of post-earthquake assessments. In recent years, terrestrial images from smartphones or cameras have become a rich source of disaster information that may be useful in assessing building damage at a lower cost. In this study, we present an efficient method of building damage detection based on terrestrial images in combination with an improved YOLOv5. We compiled a Ground-level Detection in Building Damage Assessment (GDBDA) dataset consisting of terrestrial images with annotations of damage types, including debris, collapse, spalling, and cracks. A lightweight and accurate YOLOv5 (LA-YOLOv5) model was used to optimize the detection efficiency and accuracy. In particular, a lightweight Ghost bottleneck was added to the backbone and neck modules of the YOLOv5 model, with the aim to reduce the model size. A Convolutional Block Attention Module (CBAM) was added to the backbone module to enhance the damage recognition effect. In addition, regarding the scale difference of building damage, the Bi-Directional Feature Pyramid Network (Bi-FPN) for multi-scale feature fusion was used in the neck module to aggregate features with different damage types. Moreover, depthwise separable convolution (DSCONV) was used in the neck module to further compress the parameters. Based on our GDBDA dataset, the proposed method not only achieved detection accuracy above 90% for different damage targets, but also had the smallest weight size and fastest detection speed, which improved by about 64% and 24%, respectively. The model performed well on datasets from different regions. The overall results indicate that the proposed model realizes rapid and accurate damage detection, and meets the requirement of lightweight embedding in the future.