Semi-supervised Accurate Recognition And Inverse Process Reconstruction Of Fine 3D Window Structures On Building Facades

With the continuous development of digital city and realistic 3D,high precision urban building reconstruction becomes an urgent need.As one of the most important components of the building facades,the reconstruction effect of windows directly affects the modeling effect of urban architecture.Existing automatic window reconstruction methods tend to focus only on the size and location of windows,ignoring the internal fine structure.To solve this problem,a semi-supervised identification and inverse processivity reconstruction method for fa(?)ade windows based on generative adversarial networks is proposed to address the difficulties of obtaining labeled data,ineffective utilization of unlabeled data,and significant loss of supervised learning accuracy in the current modeling process,which mainly includes the following four aspects.1.Based on the structural characteristics of windows,this paper proposes a parameterbased grammar for expressing the fine structure of windows.The grammar takes the window cell as the basic unit to grammatically describe the position,size and internal row distribution of windows.Based on this grammar,this paper classifies the window types and corresponding parameters of the existing datasets.2.Based on the window grammar and parameter settings,this paper proposes a set of process modeling process for fa(?)ade windows.The method reconstructs and combines the inner and outer borders separately to generate individual window models.Based on the distribution characteristics of building fa(?)ade windows,batch generation and result optimization are performed in groups.For the needs in practical applications,this paper also implements batch manual editing based on human interaction.3.For window image recognition,this paper implements a semi-supervised image recognition method based on generative adversarial networks.The method uses a large amount of unlabeled data and a small amount of labeled data to participate in the training of the adversarial generative network,which solves the problems of insufficient labeled data and low training accuracy in practical applications,and achieves high-precision window image classification and parameter regression based on a small number of samples.4.Based on the above research,this paper realizes the functions of network construction,model generation and result editing based on C++ and Sketchup,and completes a relatively complete set of functional building fa(?)ade window refinement reconstruction prototype system.Based on the method proposed in this paper,a variety of different buildings are selected for the fa(?)ade window reconstruction.The experimental results show that with only a small number of labeled samples for training,the method in this paper can achieve an average improvement of more than 5% in classification accuracy and parameter regression accuracy compared with the traditional supervised learning approach,and has better accuracy and visual effect in modeling results.Meanwhile,the cluster-based modeling results optimization method proposed in this paper has also been proved to be effective in experiments,which can further improve the modeling results based on the modeling results of this paper.