震损快速识别算法在湖北应城M 4.9级地震中的应用研究

Application of seismic damage rapid identification algorithm in Yingcheng M 4.9 earthquake in Hubei Province

针对基于人工检测方法存在效率低、实时性差等问题,提出了一种空间通道注意力机制改进的Faster RCNN的砌体结构震损图片快速识别算法。基于湖北省应城M 4.9级地震现场调查获取的砌体结构震害图片,制作砌体结构门窗洞口和震损类型的数据集;通过Mosaic方法对数据集进行数据增强后,构建空间通道注意力机制改进的Faster RCNN模型提取震害图片高级语义特征;使用湖北应城M 4.9级地震砌体结构震害调查数据集对模型进行训练及验证并确定最终的模型超参;最后基于改进的Faster RCNN对砌体结构门窗洞口和震损类型进行快速检测。实验结果表明,该改进的算法可以有效的识别出门、窗、剥落、裂缝,其检测精确分别为:93.1%、97.6%、74.8%、62.3%。此外,单张震害照片检测时间为60 ms,为砌体结构震害快速检测提供了新的思路。

Aiming at the problems of low efficiency and poor real-time performance based on manual detection method,a rapid detection method for masonry structure seismic damage using the improved faster region-based convolutional neural network(Faster RCNN)model modified by the spatial channel attention mechanism was proposed in this paper.Based on the seismic damage pictures of masonry structures obtained from the M 4.9 earthquake in Yingcheng,Hubei Province,a data set of basic components and common seismic damage types of masonry structures was established.After the data set was enhanced by the Mosaic data augmentation,the improved Faster RCNN model was constructed to extract the high-level semantic features of earthquake damage images.The final model hyperparameters was determined after training and verifying.Finally,the common components and seismic damage types of the masonry structures could be rapidly detected by the improved Faster RCNN model.The experimental results show that the improved Faster RCNN model can effectively detect doors,windows and walls peeling and cracks.Moreover,the detection accuracy of doors,windows and walls peeling and cracks was 93.1%,97.6%,74.8%and 62.3%,respectively.In addition,the detection time of a single earthquake damage photo is 60 ms,which provides a new perspective for rapid detection of seismic damage in masonry structures.