A Research On The Technology Of Building Segmentation And Recognition

Buildings convey the broadest information in a city.The Effective technology of building segmentation and recognition plays a significant role in academic researches and heralds vast potentials for future development.As an important part of artificial intelligence(AI),the accurate and efficient technology of building segmentation and recognition is conductive to rapidly conducting information analysis on buildings in surroundings by unmanned vehicles and unmanned aerial vehicles(UAVs),locating their own positions and fulfilling tasks,including unmanned delivery,reconnaissance and rescue.This paper carried out related work based on building images and studied the technology of building segmentation and recognition:With respect to researches on the technology of building segmentation,this paper respectively conducted researches on the segmentation technology for the edge-based buildings and watershed-based buildings and introduced the principles of two kinds of segmentation technology.The results were obtained through experiments.Consequently,this paper put forward Mask R-CNN to segment building images,analyzed the structure of Mask R-CNN,designed the flow of Mask R-CNN in the segmentation of building images and employed Label Me to make datasets of building images.Regarding researches of using support vector machines(SVM)to recognize buildings,this paper proposed an improved method of Histogram of Oriented Gradient(HOG)to extract the contour features of building images,combined HOG with Canny Edge Detector to draw eigenvectors of building images,took advantage of eigenvectors to describe contour features of building images,classified building images by SVM of V-SVR in Libsvm,selected the appropriate kernel functions through experiments,determined the parameters of kernel functions by the grid search,and tested and verified the recognition methods by building datasets of Sheffield.In the researches of adopting Convolutional Neural Networks(CNN)into building recognition,this paper proposed Mobile Net--a mobile network to recognize building images,came up with an improved structure of Mobile Net to recognize building images,employed datasets of Sheffield to verify the recognition performance of the improved Mobile Net,and compared it with the recognition performance of Inception V1 and VGG16.In order to measure and verify the recognition effect after the segmentation of building images,building datasets of class A and class B were constructed on the foundation of building datasets of Sheffield.Furthermore,they were used to test the improved method of building recognition.The experimental results showed that the average accuracy of building recognition using the improved HOG was over 97%,and the average accuracy of building recognition using the improved Mobile Net was up to 99.85%,thereby achieving favorable recognition effect.