Study Of Building Image Clustering Based On Deep Learning

Building image detection and clustering is an important task of computer vision,which is very important in the design of earthquake insurance products.Compared with traditional methods,building image detection and clustering based on deep learning has obvious advantages,such as high accuracy of detection and clustering and no need to specify artificial features,etc.In this paper,building image detection method and building image clustering method based on convolutional neural network are studied.The main work is embodied in the following two aspects:(1)Building Image Detection Based on Faster-RCNN.Several typical building images are selected as training samples to train convolutional neural network.The detection results under different sample size,different sample resolution and different sample type are compared,and the optimal parameters of Faster-RCNN are determined.In this paper,the detection accuracy of multi-class buildings can reach 85%,which basically meets the needs of practical application.(2)Building image clustering based on convolutional neural network(CNN).In this paper,different types of building images are collected as training samples to explore the application performance of convolutional neural network(CNN)in building image clustering.The over-fitting phenomenon of experimental results is improved and the discarding rate and its causes are analyzed.In this paper,the clustering accuracy of multi-class buildings can reach nearly 98%,which fully meets the actual needs.