Research On Building Fa?ade Element Extraction Based On Multidimensional Semantic Ensemble Learning

Buildings are an important carrier and expression of cities.Building fa?ade elements are an important foundation for the construction of digital cities and smart cities.Building fa?ade exhibit an array of textures and geometric forms.However,the city environment is complex.The image acquisition process is easily affected by external conditions such as foreground occlusion,shadows,and lighting.Moreover,the color and texture features of fa?ade in special scenes such as dilapidated buildings are not robust.3D point clouds have large data volume,low recognition efficiency,and lack of texture information.Therefore,how to extract building information from 2D and 3D data with high accuracy and efficiency has also become a key problem for research in various fields.This paper studies the extraction method of building elements by ensemble learning multi-dimensional semantic features.The main research contents and conclusions are as follows:(1)A method for constructing a multi-dimensional semantic feature projection correlation map is studied.Building facades contain both 2D color texture features and 3D spatial structure features.Based on 2D or 3D data alone,it is difficult to accurately characterize the morphology of building facades.To address this issue,a method for constructing a projection correlation map that integrates multi-dimensional semantic features is proposed.Using multi-view image sequences as the data source,the dense point cloud model of the building is reconstructed through motion recovery structure theory and multi-view stereo vision algorithm,and the building’s 3D orientation and 3D curvature features are calculated.Based on the physical-object relationship model,a multi-dimensional semantic feature projection correlation map is obtained through backprojection.Experimental results show that the multi-dimensional semantic correlation map characterizes the spatial structure features of the building,effectively distinguishes different facade areas with pixel textures,and the accuracy of the correlation map backprojection is within 1 pixel,achieving pixel-level fusion of 2D and 3D features of building facades.(2)A superpixel segmentation method based on multi-feature semantic integration learning is studied.Image-based segmentation methods are not suitable for building facades with complex spatial forms,while point cloud-based segmentation methods have low efficiency and are difficult to consider multi-dimensional features.To address these issues,a superpixel segmentation method that integrates multi-feature semantics through ensemble learning is proposed.Using a fully convolutional neural network as the base learner,a superpixel segmentation result that considers multi-semantic features is obtained by predicting and fusing soft association mappings based on the projection correlation map.Experimental results show that the superpixel segmentation based on multi-feature semantic integration learning is superior to image-based segmentation methods in terms of segmentation accuracy,edge adherence,and better fits the real building model,while also balancing efficiency and accuracy.(3)A building element extraction method based on posture semantic density analysis is studied.Building morphology is complex and diverse,and traditional image clustering methods are easily affected by texture and color interference,and it is difficult to consider spatial form information,resulting in poor clustering results.To address these issues,a posture semantic density analysis-based superpixel hierarchical clustering and merging method is proposed.The posture semantic density distribution graph is constructed using the multi-feature semantic superpixel segmentation result as nodes.Combined with the multi-dimensional semantic correlation map,the hierarchical clustering and merging strategy is implemented to achieve fine extraction of building facade elements.Experimental results show that the building facade elements extracted by the proposed method have a Pixel Accuracy(PA)rate of over 70%,and the highest mean Intersection over Union(m Io U)is 95%,which is better than the method based solely on image clustering.The proposed method takes into account multi-dimensional semantic features and effectively improves the accuracy of building facade element extraction.