基于生成对抗网络的建筑物损毁检测EI北大核心CSCD

建筑物作为自然灾害中最受影响的承灾体之一,其损毁信息的准确提取对灾后应急救援具有十分重要的意义。本文借鉴多模态的思想,提出了一种自动检测损毁建筑物的recursive-generative adversarial networks(RS-GAN)方法,将损毁建筑物检测分为灾前建筑物识别和灾后损毁建筑物检测两个任务,且分别在两个GAN分支中完成。RS-GAN加入联合损失函数将两个GAN分支进行连接,充分利用两个任务之间的潜在互利性提升检测效果。RS-GAN利用第1条GAN分支识别建筑物灾前形状与位置,并将识别结果作为第2条GAN分支的输入进行损毁建筑物检测任务,从而使检测结果具有更清晰的轮廓。该方法为端到端模型,在不需要过多的人工干预情形下,实现了损毁建筑物的自动检测。为了验证RS-GAN模型的效果,在圣罗莎和密苏里两个数据集上进行了测试。试验结果表明,RS-GAN方法拥有更好的检测性能,在圣罗莎数据集上的总体精度和平均精度分别达到了0.90和0.86。

Detection of damaged buildings based on generative adversarial networks

As one of the most affected hazard-affected bodies in natural disasters,accurate damage information extraction of buildings plays a significant role in post-disaster emergency rescue.Referring to the idea of multi-mode fusion technique,a recursive generative adversarial networks(RS-GAN)method is proposed to automatically detect damaged buildings.In RS-GAN,the workflow of damaged buildings detection is composed of two sub-tasks as follows:building identification before disasters as well as damaged building detection after disasters,which are completed in two GAN branches respectively.Specifically,RS-GAN adds a joint loss function to connect the two GAN branches,making full use of the potential mutual benefit between the two tasks to improve the detection accuracy.In addition,the results of building identification are added to the damaged building detection task to locate potential damaged areas.The method proposed in this paper is an end-to-end model,which can automatically detect damaged buildings without excessive manual intervention.To verify the effect of the RS-GAN model,in this paper,two experiments were set with the Santa Rosa dataset and Missouri respectively.Experimental results show that RS-GAN method has better detection performance compared to other competitive methods,and the overall accuracy and average accuracy on the Santa Rosa data set are 0.90 and 0.86,respectively.