An Improved Instance Segmentation Method for Fast Assessment of Damaged Buildings Based on Post-Earthquake UAV Images

Area of Science:

• Computer Vision
• Remote Sensing
• Disaster Management

Background:

• Accurate building damage assessment post-disaster is crucial for emergency response.
• Existing semantic segmentation and object detection methods struggle with high-resolution UAV imagery, facing challenges in multi-damage categories within buildings and precise edge extraction.
• These limitations hinder effective post-disaster rescue and fine-grained damage evaluation.

Purpose of the Study:

• To develop an improved instance segmentation model for accurate and efficient fine-grained building damage assessment from UAV imagery.
• To enhance classification accuracy and improve small object segmentation for better damage evaluation.
• To address the limitations of current methods in handling complex damage scenarios and building edge extraction.

Main Methods:

• Proposed an improved instance segmentation model incorporating a Mixed Local Channel Attention (MLCA) mechanism in the backbone to boost classification accuracy.
• Refined the Neck part of the model to enhance the segmentation accuracy of small objects.
• Tested the model on Unmanned Aerial Vehicle (UAV) images from the Yangbi earthquake and compared it with state-of-the-art models like Mask-R-CNN and YOLO V9-Seg.

Main Results:

• The modified model demonstrated superior performance, outperforming the original model by 1.07% in mAPbbox50 and 1.11% in mAPseg50.
• Significant improvements were observed in classification accuracy for intact (2.73%) and collapse (2.58% and 2.14%) categories.
• The proposed model achieved comparable accuracy to state-of-the-art methods while being three times faster, showcasing enhanced efficiency.

Conclusions:

• The improved instance segmentation model offers a valuable solution for fine-grained building damage evaluation using UAV imagery.
• The integration of MLCA and Neck refinement effectively addresses challenges in damage classification and small object segmentation.
• The model's enhanced accuracy and efficiency provide a significant advantage for post-disaster emergency response and assessment.