Research On Building Segmentation Of Aerial Images Based On Convolutional Neural Network

With the development of social economy,the pace of urbanization in China is speeding up day by day,and the number and category of urban buildings are also growing rapidly.Digital management puts forward higher requirements for 3D reconstruction of urban buildings.In addition,with the popularization of UAV aerial photography technology,the aerial images of urban buildings are easier to obtain.These aerial images contain rich building feature information.It is of great significance to detect and segment building objects from the aerial images for 3D reconstruction of cities.At present,the common image segmentation methods include threshold segmentation and edge detection,but these methods have some limitations.It is difficult to achieve accurate segmentation results in the complex high-resolution aerial image building segmentation task.Based on the superior performance of convolutional neural network algorithm in the field of computer vision,aiming at the problems of the traditional image segmentation method in the aerial image building segmentation scenario,this paper applies convolutional neural network algorithm to the aerial image building segmentation task,the goal is to segment the building contour area needed for building 3D reconstruction from the aerial image with high precision.The main work of this paper is summarized as follows.Aiming at the problems of high cost of obtaining the top view plan of the building's main body contour,low precision of aerial image segmentation and interference of building roof,DEXTR(Deep Extreme Cut)which is a semi-automatic image segmentation method is improved,and a new method E-DEXTR(Enhanced Deep Extreme Cut)which is based on the deep residual network,is proposed to represent the positions of five points as additional input channels of the network.The goal is to achieve the accurate segmentation of the buildings in the aerial images.The experimental results show that compared with the traditional semi-automatic segmentation method Grabcut,E-DEXTR method has higher segmentation accuracy and efficiency in aerial image building segmentation task;compared with DEXTR method,E-DEXTR method has better robustness and anti-interference.The proposed E-DEXTR method can provide high-precision building top profile and building top picture for building 3D reconstruction task.It can also be used in the production process of aerial image building data set as an accurate and effective mask annotation tool or semi-automatic contour annotation tool to improve the efficiency of data set annotation.For large-scale building segmentation scenario,aiming at the problems of low accuracy and high cost of building detection and segmentation in aerial images,a method of building detection and segmentation based on Mask R-CNN is proposed to realize the automatic and efficient detection and segmentation of buildings in aerial images.In the method,aiming at the problem of poor detection and segmentation effect of the trained model in aerial images with shadow and other interferences,a non random masking data augmentation method is proposed to improve the segmentation accuracy of the model for buildings with shadow and other interferences.In addition,the effects of the use of transfer learning method,data augmentation method,selection of different depth of residual network in the feature extractor module on building detection and segmentation model are tested by experiments,and the generalization performance of the model is tested on satellite remote sensing data.Compared with the E-DEXTR method proposed in this paper,which needs to label five points manually to accurately segment the building target,this method can automatically detect the target area and segment the building.In summary,this paper solves the problems of low segmentation accuracy,long time and high cost in aerial image segmentation task by studying the application of convolutional neural network in the field of image segmentation.The two methods proposed in this paper have their own advantages.Different methods can be selected according to different requirements of building 3D modeling.The E-DEXTR method is simple in operation and high in segmentation precision,which is suitable for the scenario with few buildings to be segmented and high in segmentation precision;the method of building detection and segmentation in the aerial images based on Mask R-CNN can automatically detect and segment buildings,and can segment a large number of buildings in the aerial image in batches,which is suitable for the scenario with a large number of buildings to be segmented and high efficiency requirements for segmentation.