Research On Building Damage Degree Model Based On Mask RCNN

China is a country prone to natural disasters.In the face of sudden natural disasters,it is very important for the follow-up post disaster rescue and reconstruction to grasp the disaster situation quickly and quantitatively assess the post disaster losses in time.The rapid development of image processing technology and deep learning provides a new idea for post disaster loss assessment.In this thesis,an improved image stitching algorithm is proposed,and a building disaster index model based on Mask RCNN is established,so as to quickly obtain the post disaster panoramic aerial image of the disaster area and the comprehensive quantitative assessment of post disaster loss.Aiming at the acquisition of panorama after disaster,this thesis proposes an improved fast image stitching algorithm based on ORB to improve the image stitching speed without reducing the accuracy.In this thesis,by comparing the image stitching speed of image stitching algorithms based on SIFT,SURF and ORB features,it is found that the main time consumption of image stitching comes from the feature extraction part.Combined with the shooting path of UAV aerial images,the returned aerial images have the characteristics of pixel overlap at the top,bottom,left and right.This thesis proposes an improved region based feature extraction algorithm.The experimental results show that the application of this algorithm to the image stitching algorithm based on SIFT,SURF and ORB features can effectively improve the image stitching speed of the three image stitching algorithms,and almost no loss of accuracy.The more complex the feature extraction algorithm is,the more obvious the improvement effect is,especially with the increase of the number of Stitching image,the time consumption advantage of this algorithm is more obvious.The improved image stitching algorithm based on ORB in this thesis only takes 15.25 seconds to realize the image stitching of 12 images with a resolution of 5964×5304 panoramic image stitching,compared with before the improvement saves 1/5 of the time.For the quantitative evaluation of post disaster losses,according to the distribution of disaster affected bodies in natural disasters and the realizability of data statistics,this thesis puts forward four evaluation indexes of building damage degree,life loss degree,disaster area and economic loss degree,and establishes the index model of building disaster degree based on the evaluation indexes,Finally,a case study of Lushan County,Ya’an City,Sichuan Province is conducted to verify the effectiveness of the model.In order to get the degree of building damage,this thesis trained the adjusted positioning and classification model based on UNet and Resnet50,semantic segmentation model based on Deeplab V3+,and instance segmentation model based on Mask RCNN on x BD data sets,and compared their performance on the verification set.It was found that the instance segmentation model based on Mask RCNN achieved the best₁ score on this datasets,and in comparison with the existing models,it also achieves the best,so this thesis finally selects the building damage evaluation model based on Mask RCNN to evaluate the degree of building damage.In order to improve the prediction effect of the model on the aerial images,the post earthquake aerial images of Wenchuan,Changning and Lushan in Sichuan are annotated,and the post disaster aerial datasets are established.By further training the pre trained model on the aerial data set,the prediction accuracy of the model is improved,and the transitivity of the neural network in building damage assessment is verified.