Research On Texture Segmentation And Structure Primitive Matching Method For Object Stereo Reconstruction

With the development of remote sensing satellite sensor technology,people can obtain traditional data products including spectral images,optical images through satellite,aviation and drone platforms.At the same time,the stereo information combining the optical and structural attributes also can be acquired by remote sensing information processing technology.Remote sensing object stereo information breaks the cognition of traditional remote sensing information.By using the combination of object surface optical physical attribute information and object three-dimensional structural information,it can represente the essential attributes of the object more sufficiently in real space.In recent years,with the popularity of digital information technology,the object three-dimensional construction have been widely applied in various field related to people's livelihood.Further exploration of the application potential of the object stereo construction will greatly enhance effectiveness of implementation for many fields,including urban planning,geographic exploration,earthquake disaster reduction,as well as military reconnaissance.They all have major significance for the development of national economy and national defense.The traditional three-dimensional construction methods usually aim at constructing large area terrain scenes,and use the principle of photogrammetry to construct the object structure information in a wide range of scenes from the satellite image by means of geometric feature extraction and stereo matching,which often ignore the analysis of the properties of the object surface optical information.In addition,the traditional three-dimensional construction methods have more high requirements for data accuracy and low universality,and it is difficult to ensure the structural integrity and construction accuracy of the three-dimensional structure for a complex object.Different from the traditional three-dimensional construction methods,the reconstruction method of remote sensing objects proposed in this paper is researched deeply from the view of the object surface texture information segmentation and the overall structure construction of objects,so as to obtain the remote sensing object stereo construction results with rich object texture information and complete structural information.The method proposed in this paper is focusing on the object-driven stereo construction strategy.The main research content includes three parts:The first part of the research content conducts the object surface texture information segmentation by analyzing the object texture differences based on structural attributes.The second part mainly studies the primitive structural feature matching methods.The purpose is to select the predefined primitive topology connection structure from the model library.On this basis,the third part presents the parametric representation of the overall structure of the object through the overall structural constraints of the primitive topology,and combines the method of parameter optimization to achieve the goal of constructing the overall three-dimensional structure.Finally,combining the extracted object surface texture information and the object three-dimensional structure information,the stereo construction results of the remote sensing object are obtained.According to the summary of research contents of remote sensing object stereo construction,the three-dimensional construction method proposed in this paper mainly consists of three key technologies: object surface texture segmentation method,object primitive structure topology matching method,and the optimal parameter estimation method of object overall structure.The following three issues have conducted deeply research on remote sensing object stereo construction.First of all,due to the high homogeneity of texture information in the remote sensing object region and the small difference in texture distribution within the object,it is impossible to accurately obtain the texture information of each surface of the object.To solve this problem,this paper proposes a texture segmentation method based on low-rank and sparse cartography.The method uses the object-based image analysis method(OBIA)as a theoretical guide.By analyzing the statistical features of the object's local textures,a related cartography that can characterize the structural differences of the object texture in the feature space is constructed to overcome the influences of the noise and singularity intraclass texture segmentation.In the aspect of features extraction,a local texture histogram feature extraction method is proposed,which can effectively improve the description ability of the object local texture features.In order to further analyze the difference in the structural distribution of local textures,a correlation map based on low rank and sparse characteristics is constructed in the feature space,which analyzes the structural differences of the object texture information from the perspective of accurately captures the structural distribution differences of object inner textures with high homogeneity.In the end,the texture segmentation results of each object surface can be obtained by integrating with the graph cut method.Experimental results verifies that the texture segmentation algorithm proposed in this paper can accurately obtain the texture information of each object surface and provide important support for obtaining complete stereo construction results.Then,in the process of object primitive structure matching,due to the reason of insufficient description ability of shape feature of objects,which results in the problem that the primitive structure can not be accurately matched from the predefined model library.Therefore,an object primitive topology matching method based on deep embedded network is proposed.In this method,different kinds of typital primitive structure are predefined in the parametric model library from the global and local perspective of object structural characteristics.Based on the primitive library,an object primitive topology matching algorithm based on deep embedded network is proposed.The object shape features can effectively identify the differences within the object class extracted by the deep embedded network learning.Then,it uses feature matching technology to match the object primitive topology with complete and normalized structural attributes from the model library,which provides global topology structure constraints between primitives for the final object stereo model construction and effectively guarantees the structural integrity and standardization of the object model stereo construction.The experimental results verifie that the method proposed in this paper can accurately match the primitive topology structurein the model library,which provides important global structural guidance for further obtaining high precision and complete object three-dimensional construction.Finally,in order to overcome the problem of the low precision of object overall structure reconstruction,which is caused by the uncomplete description of object structure.In this paper,a parameter estimation algorithm is proposed for object reconstruction based on the topological constraints of prmitive objects.To fully describe global and local structure of the object,a structural parametric representation model based on object primitive topological constraints is proposed.On one hand,the model function can represent the top three-dimensional structure of the object primitives with different complexity.On the other hand,by introducing the topoligical constraints into the function of object model,this method can effiectively insure the global structural characteristics of the object structure,and also provide the guarantees for completely constructing the stereo model of object.In the aspect of parameter optimization estimation,against the multi-parameter optimization problem of the object's overall structure,the particle swarm optimization algorithm with simple rules,high accuracy and fast convergence is used to estimate the optimization parameters of the overall structure of the object,and the object model stereo construction results can be acquired with complete structure and high accuracy.The experimental part verifies that the method proposed in this paper can complete the construction of the object top structure and ensure the precision of the model construction.