Research On Acquisition Method Of Building Attributes Based On Convolution Neural Network

The classification and statistical data of buildings in cities are important to the design of earthquake insurance products.Rapid and accurate acquisition of those data is critical to set the parameters of the insurance of the products.Obtaining the building structural attributes from street view data has significant advantages over traditional methods,however,how to obtain these structural characteristics accurately still pose challenges.This paper studies the method of acquisition building attributes based on convolutional neural network,which realizes the effective acquisition of building's attributes from images.The major works are as follows:(1)Fisheye-images correction.The five correction methods are implemented,including: longitude based correction method,latitude and longitude based correction method,spherical projection based correction method,polynomial based correction method,and division based correction method;and the image distortion correction experiments on the same building fisheye-images are carried out.The experimental results indicate that the correction method based on spherical projection is better than the other four correction methods.Therefore,the spherical projection based correction method is chosen to correct the distorted street view images.(2)Floor numbers estimation.The VGG16 convolutional neural network is selected as the basic network framework and the transfer learning is applied.Four factors that critical to the parameters of VGG16 are considered,including: image numbers,the image size,the building street view images with distortion and the street view images of building with obstructions.Experimental comparisons are carried out and the optimal model parameters are obtained.Finally,the model of estimating the number of floors is determined.