眭海刚, 黄立洪, 刘超贤. 利用具有注意力的Mask R-CNN检测震害建筑物立面损毁[J]. 武汉大学学报 ( 信息科学版), 2020, 45(11): 1660-1668. DOI: 10.13203/j.whugis20200158
引用本文: 眭海刚, 黄立洪, 刘超贤. 利用具有注意力的Mask R-CNN检测震害建筑物立面损毁[J]. 武汉大学学报 ( 信息科学版), 2020, 45(11): 1660-1668. DOI: 10.13203/j.whugis20200158
SUI Haigang, HUANG Lihong, LIU Chaoxian. Detecting Building Façade Damage Caused by Earthquake Using CBAM-Improved Mask R-CNN[J]. Geomatics and Information Science of Wuhan University, 2020, 45(11): 1660-1668. DOI: 10.13203/j.whugis20200158
Citation: SUI Haigang, HUANG Lihong, LIU Chaoxian. Detecting Building Façade Damage Caused by Earthquake Using CBAM-Improved Mask R-CNN[J]. Geomatics and Information Science of Wuhan University, 2020, 45(11): 1660-1668. DOI: 10.13203/j.whugis20200158

利用具有注意力的Mask R-CNN检测震害建筑物立面损毁

Detecting Building Façade Damage Caused by Earthquake Using CBAM-Improved Mask R-CNN

  • 摘要: 震后建筑物损毁信息是灾情快速评估和应急救援的重要决策依据之一。针对传统的建筑物损毁遥感检测技术只关注于顶面信息,导致众多顶面结构完好而中间层、底层倒塌或崩裂的损毁建筑物处于检测盲点的问题,提出了一种融合引入注意力机制的深度学习实例分割模型和图像多尺度分割算法进行震后建筑物立面损毁检测的方法。首先,利用卷积块状注意力模块(convolutional block attention module, CBAM)改进掩模区域卷积神经网络(mask region-based convolutional networks, Mask R-CNN)模型实现了复杂建筑物立面背景中的损毁信息提取;然后,基于建筑物立面影像多尺度分割结果,利用多数投票规则实现了损毁检测结果的后处理优化。实验结果表明,相比传统损毁检测方法,所提方法能够更有效地实现震后建筑物立面损毁信息的精准定位,总体准确率可达到89.15%。

     

    Abstract:
      Objectives  Building damage information can provide an important basis for the decision making of rapid post-earthquake assessment. Traditional building damage detection techniques mainly focus on the roof surface, thus many damaged buildings with an intact roof surface but collapsed middle floors may be neglected. We propose a method of building façade damage detection based on deep learning and multiresolution segmentation algorithm.
      Methods  The method which integrates the instance segmentation and multiresolution segmentation algorithm is applied to detecting the post-earthquake building façade damage. The first thing is to collect the ground images of post-earthquake buildings in the field and perform the data augmentation. Secondly, we use the convolutional block attention module (CBAM) to improve Mask R-CNN. Then the dataset is input to the improved model for training, and finally a multiresolution segmentation algorithm is adopted to post-process the building façade damage detection results output by the CBAM-Improved Mask R-CNN.
      Results  The experimental results show: (1) Collecting ground images of buildings in the field and performing image augmentation can effectively guarantee the necessary training sample size of the instance segmentation model. (2) The Mask R-CNN improved by CBAM attention mechanism significantly improves the post-earthquake building facade damage detection capabilities, which realizes the precise extraction of damage information from complex building façade backgrounds. (3) In addition, using the multiresolution segmentation algorithm to post-process the building facade damage detection results can obviously solve the blurred boundary problems caused by the accumulation of convolutional layers.
      Conclusions  The proposed method can significantly improve the capability of post-earthquake building façade damage detection when compared to the traditional methods, which also raises the Mask R-CNN's accuracy, precision, recall and F2-score to a certain degree. It can be inferred that the proposed method has the strong potential to be applied to the post-earthquake building façade damage detection and therefore provides an important technical means for detecting the comprehensive and detailed building damage detection caused by earthquake.

     

/

返回文章
返回