Building Extraction Using The Fusion Of Point Cloud And Image Base On Deep Learning Method

With the rapid development of remote sensing technology,the acquisition of data sources has become easier.High-resolution remote sensing images have rich texture information and spectral information,which provide favorable features for the identification and extraction of buildings,and are currently widely used in the field of building extraction and recognition.However,it cannot provide the specific elevation information of the building,and cannot completely separate the building from the ground to obtain high accuracy.Therefore,combining remote sensing images and point clouds to produce joint image extraction becomes more effective and feasible.Conventional methods are often based on manual feature setting.This method cannot learn the deeper semantic features of buildings,so the extraction results of complex buildings are unsatisfactory.In order to solve the shortcomings of conventional methods,this paper summarizes the shortcomings of existing extraction methods,and on this basis,conducts research on deep learning.The main research contents and results of the joint building of point cloud and image based on deep learning are as follows:1.In-depth explanation of the basic principles of deep learning convolutional networks,and then in-depth study of building extraction methods based on deep learning methods,and analysis of the causes of adverse effects on the accuracy results.2.Aiming at the current problems of inaccurate building boundary extraction by deep learning convolutional network and errors and omissions of small buildings,it is proposed to use the attention mechanism to drive the joint image for building extraction.This method mainly uses the channel attention mechanism to control the channel.The effective semantic feature enhancement and spatial attention mechanism enhance the effective semantic feature of spatial location,and effectively improve the accuracy of building extraction.3.In order to verify the effectiveness of the method in this article,the above method strategy is applied to two open source data sets,and three evaluation indicators of intersection ratio(Io U),F1?Score(F1),and pixel accuracy(OA)are used for accuracy evaluation..Among them,the best results in Igarss data are Io U 0.8473,F10.8606,and OA 0.9489.The best results in Potsdam data are Io U 0.9291,F1 0.9491,and OA 0.9684,which are significantly higher than the extraction accuracy of traditional methods.Then,in order to prove the generalization of the strategy of this method,the multi-view oblique joint image of Wuhan area was extracted,and the method achieved good results of Io U 0.8414 and F1 0.8920 OA 0.9188 on the oblique image.Therefore,it can be proved that the method strategy has good generalization and robustness.