基于深度学习的建筑破坏状态智能评估研究

Research on intelligent evaluation of building damage state based on deep learning

地震发生后需要对震害区域建筑物破坏等级进行快速评估,以便于震后辅助决策和应急救援。通过分析建筑物外立面震害照片数据,挖掘不同建筑破坏等级与其立面图像特征潜在的映射关系,建立基于震害图像数据的建筑破坏状态智能评估深度学习方法,并应用于都江堰的地震调查数据。首先,训练获得了DeeplabV3+图像语义分割深度神经网络模型,实现在震后复杂背景下的建筑物外观整体的分割提取。进一步,利用迁移学习进行ResNeXt网络参数训练用于图像分类,实现建筑物地震破坏等级的评估。分析了都江堰地震现场调查数据,实验结果表明,所提方法能够较精细地将建筑破坏划分为基本完好、损坏和倒塌三类,准确率达到90.33%。评估模型可直接应用于震后建筑状态的较精细评估,且对外立面图像拍摄角度无较高要求,对图像背景鲁棒,可避免环境因素影响。相较于人工判断,研究方法提高了工作效率,缓解了调查者主观性判断带来的离散性影响,具有良好的应用潜力。

After an earthquake,it is necessary to quickly assess the building damage state in the earthquake-damaged area,so as to assist the decision-making and emergency rescue.Considering the potential mapping relationship between buildings with different damage levels and their facade features based on facade photos of buildings,an intelligent assessment method of building damage state based on earthquake damage photos is established.This method is applied to the survey data of Dujiangyan in the Wenchuan earthquake.Firstly,the segmentation and extraction of the overall appearance of buildings in complex post-earthquake backgrounds is achieved by a trained image semantic segmentation neural network model DeeplabV3+.Furthermore,ResNeXt network parameters for image classification are obtained by transfer learning.Then the model is used to realize the damage state assessment of buildings.The experimental results show that the proposed method can finely assess the damage state of buildings with the accuracy of 90.33%,and automatically classify them into three categories that are basically intact,damaged and collapsed.After the training,the model can be directly used for fine-grained assessment of the post-earthquake building state,and there is no high requirement for the image capture angle.It can achieve the effect of being robust to the image background and avoiding the influence of environmental factors.Compared with manual judgment,it improves the efficiency,avoids the subjective judgment of investigators,and has good application prospects.