Scholarlink

Disasters are inevitable and can severely affect the infrastructure. Infrastructure monitoring must be automated to reduce losses and expedite rescue efforts. To expedite emergency response and resource allocation, automating the post-disaster assessment of building damage is crucial. Deep learning has been widely used in computer vision in recent years, and substantial progress has been made in this field. This paper presents the deep learning approach for building damage level detection using YOLO, a model known for its real-time object detection capabilities. Data required for training and testing is acquired from Roboflow and Kaggle. The acquired dataset is annotated considering damage in three classes: minor cracks, major cracks, and collapsed buildings. Findings are presented using Yolov8 Medium and Yolov8 Nano implementation, tested on customized annotated buildings dataset. Experimental results prove that the Yolov8 Medium outperforms as compared to the Yolov8 Nano model. Findings also demonstrate that faster R-CNN gives better accuracy than Yolov8. Findings are presented with evaluation results. Performance of each model noted as Faster R-CNN (with RESNET-50): Validation Accuracy: 82%, YOLOv8 Medium: Validation Accuracy: 54% and YOLOv8 Nano: Validation Accuracy: 52%

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performance Analysis of a Deep Learning-Based Object Detection Approach for Post-Disaster Building Damage Level Assessment Using YOLO and Faster R-CNN

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Shilpa Pimpalkar,
  • D. Sai Madhavi,
  • N Kushal Rao

摘要

Disasters are inevitable and can severely affect the infrastructure. Infrastructure monitoring must be automated to reduce losses and expedite rescue efforts. To expedite emergency response and resource allocation, automating the post-disaster assessment of building damage is crucial. Deep learning has been widely used in computer vision in recent years, and substantial progress has been made in this field. This paper presents the deep learning approach for building damage level detection using YOLO, a model known for its real-time object detection capabilities. Data required for training and testing is acquired from Roboflow and Kaggle. The acquired dataset is annotated considering damage in three classes: minor cracks, major cracks, and collapsed buildings. Findings are presented using Yolov8 Medium and Yolov8 Nano implementation, tested on customized annotated buildings dataset. Experimental results prove that the Yolov8 Medium outperforms as compared to the Yolov8 Nano model. Findings also demonstrate that faster R-CNN gives better accuracy than Yolov8. Findings are presented with evaluation results. Performance of each model noted as Faster R-CNN (with RESNET-50): Validation Accuracy: 82%, YOLOv8 Medium: Validation Accuracy: 54% and YOLOv8 Nano: Validation Accuracy: 52%