基于多任务学习的建筑毁伤评估方法

Multi-task learning based building damage assessment method

建筑毁伤评估在灾害救援过程中对制定救援策略、优化资源调度等方面具有重要作用。现阶段,基于语义分割的毁伤评估方法难以提取毁伤建筑细粒度语义信息,对此提出一种基于多任务学习的建筑毁伤评估方法,将毁伤评估分为粗粒度的建筑区域提取与细粒度的毁伤分割两个子任务,通过共用编解码器,与上下文信息融合模块实现建筑区域的粗粒度提取和建筑毁伤的细粒度分割,将两个结果使用Hadamard积融合,得到最终评估结果。实验结果表明,所提的基于多任务学习的建筑毁伤评估方法有较好的性能。

Building damage assessment plays an important role in the disaster relief process,influencing the formulation of rescue strategies and optimization of resource allocation.Currently,damage assessment methods based on semantic segmentation face challenges in extracting fine-grained semantic information for damaged buildings.Thus,a multi-task learning based approach for building damage assessment is proposed,dividing the damage assessment into two subtasks as coarse-grained building area extraction and fine-grained damage segmentation.The proposed method utilizes a shared encoder-decoder and context fusion module to achieve coarse-grained extraction of building areas and fine-grained segmentation of building damage.The results of these two tasks are fused using the Hadamard product to obtain the final assessment.Experimental results demonstrate that the proposed multi-task learning based building damage assessment method performs well.