Research On Building Shape Matching Methods Based On Deep Learning

With the rapid development of computer technology,shape matching technology has become a research hotspot in computer vision and other fields.Many shape matching methods have been successfully applied to face recognition,text recognition and other issues.At the same time,shape matching is also one of the important factors in cognitive recognition of geographical features.How to measure the similarity between two object shapes depends on the characteristics of the object ontology and the subjective cognition of the audience,which is a process of spatial cognition.Building elements are important components of geospatial elements,with obvious visual features such as right-angled turning and axis symmetry.The shape matching is an important basis in the issues of comprehensive simplification and data fusion.The basic process of shape matching includes extracting a certain shape representation factor and a similarity measure under the shape representation.Existing shape matching methods are mostly described based on features of certain aspects of the shape,without comprehensively considering the relationship between shape representations.From the cognitive perspective,there is no reasonable explanation for the fusion of different cognitive perspective and numerical scale shape representations by artificially defining weights.Deep learning techniques have made significant achievements in areas such as computer vision,natural language processing,and machine intelligence that rely on human cognition.Deep learning can mine cognitive features based on human data experience.Therefore,starting from spatial cognition,this paper constructs a series of building shape representation features from different cognitive perspectives.Based on feature mining and knowledge discovery ability of deep learning,the building shape is coded and represented,so as to achieve shape matching in line with visual cognition.Specifically,the main achievements and innovations of this paper include:(1)Shape representation features of different cognitive perspective are constructed.On the basis of considering the ontology features and cognitive factors of the building shape,a group of shape representation factors of building surface elements are constructed,including the global features of the shape regions and the multi-scale sequence features based on the shape contours.The features conform to the translation,scale and rotation invariance,depicting building shapes from different cognitive perspectives.(2)Constructing a building shape coding model based on deep learning.Based on the idea of seq2 seq framework and self-encoder,the model performs nonlinear mapping and dimension reduction on the basic features of the shape in an unsupervised manner,and reconstructs the original feature set to realize the shape coding ability.Through the model,a shape encoding with different cognitive perspective is obtained,which is a further abstraction of the original shape representation,and the shape ontology and cognitive information are represented by lower dimensions.(3)Propose the corresponding shape similarity calculation method and verify the analysis.Based on the comprehensive shape coding library obtained by the above method,the cosine similarity is used as the reference of the similarity measure,and the rotation sampling problem of the shape contour feature is considered,and the similarity calculation of the shape matching is performed.Finally,the scenes of simulation data and real data are designed.The validity of shape coding is verified by similarity visual analysis,shape retrieval and matching experiments.The results show that the shape coding model constructed in this paper can construct shape coding with cognitive significance.The code is a good representation of the global and local features of the shape of the building features,which is in line with visual cognition.It can be used to measure the similarity between shapes,and has sufficient distinguishing ability for different shapes,and complete the shape retrieval and matching tasks.Based on the shape representation of different cognitive perspective,the deep learning model can mine the essential features and implement feature fusion and shape coding.At the same time,there are still some problems in the method of this paper.Specifically,the model relies on multi-scale features of shape contours,and there may be problems of overfitting local contours according to different scales.In general,the building shape matching method of this paper can complete the identification and matching tasks of building shapes,and has certain reference value for deep learning in geospatial cognitive mining.