Research And Application Of Geometric Feature Detection For 3D Graphics

With the increasing popularity of 3D data acquisition equipment and the research on modeling methods,3D data representation and processing are more and more widely used in areas such as simulation,CAD design and digital twins.Geometric feature detection of 3D graphics is one of the key issues in 3D representation and processing.How to use the attributes of 3D graphics(such as:position,color,normal,etc.)to automatically extract geometric features that reflect the inherent characteristics or structural information of the 3D graphics is the core research problem.Therefore,we mainly focus on the detection and application of geometric features of 3D graphics,including sparse structured feature detection and intrinsic feature detection.For 3D graphics,sparse and concise structured features can better represent different types of objects.Moreover,the problem of high time complexity of related data processing algorithms due to the huge amount of data can be avoided.Especially for 3D point cloud data,how to extract sparse geometric features(such as keypoints)in 3D point cloud is a challenging problem due to some factors,such as data noise and resolution changes.To this end,we propose a multi-task joint learning method,which can accurately detect the location and semantic information of keypoints in the 3D model.In addition,we combine prior medical knowledge with sparse geometric feature detection,and proposes a dense coding-based tooth anatomical feature detection method.Intrinsic geometric features can represent the local geometric properties of 3D graphics.The cross field is used to represent the intrinsic characteristics of 3D data.A high-quality cross field requires that it can align the boundary features and smoothness.Therefore,it is a challenging problem to generate high-quality cross fields under complex feature constraints.We propose a method to detect cross fields on low-quality point cloud,and realize the point cloud upsampling task based on cross field optimization.Furthermore,we further apply the characteristics of the cross field to 2D image vectorization,and proposes a cross field-oriented image quadrilateral meshing method.The details are as follows:(1)3D keypoint detection is a fundamental problem in computer graphics and computer vision,especially for shape analysis and model matching.Therefore,keypoint detection is a core research problem.we propose a multi-task joint learning network for 3D keypoint saliency and correspondence estimation.To better capture the local and global features of 3D models,we design a spatial multi-scale perception module to connect feature maps of different scales in the process of point cloud feature extraction.In the multi-task joint learning process,the offset vector of each point to the keypoints in the 3D model is obtained through a voting mechanism.We also predict the confidence value of each point in the 3D model,and then filters out the low-confidence points to generate reliable voting results,so as to achieve high-precision 3D keypoint detection.(2)A core problem in the field of digital orthodontics is the extraction of anatomical features from 3D tooth models.Due to the complex geometric definition of anatomical features and the variability between different types of teeth,in most current orthodontic platforms,tooth feature points and axes need to be annotated by experienced dentists.Therefore,we propose a tooth feature detection method based on dense encoding,which encodes the feature points and axes as a geodesic distance field defined on the tooth model surface and a projected vector field from the tooth surface points to the axis points.The proposed method has been evaluated on tooth models by experienced dentists,and the detection results can reach a doctor-satisfying accuracy.(3)Compared with 3D mesh,point cloud is an unordered and irregular collection of spatial points.The point cloud acquired by 3D scanning equipment are often sparse,nonuniform and have holes,resulting in incomplete original geometric features.How to compute the cross field on low-quality point cloud data is an extremely challenging problem.We propose a method for predicting the cross field of each point cloud using a neural network.Considering the properties of the cross field,we propose a point cloud upsampling method based on iterative optimization of the cross field.The network is fed with low-quality point cloud data and learns to predict the cross and normal fields.Arbitrary magnification point cloud upsampling is achieved under the guidance of the cross field and the normal field.And through iterative optimization,the uniformity of the point cloud and the accuracy of the cross field are gradually improved,and then the upsampling results of the point cloud are further improved.(4)A cross-field-oriented image quadrilateral meshing method is proposed,and applications such as image vectorization and image editing are realized through the quadrilateral mesh.Based on the traditional 3D mesh model parameterization,the 3D cross field is introduced into the 2D image,and the cross field is initialized according to the color gradient of each pixel.The energy function of the soft constraint term is constructed and optimized to achieve smoothing of the cross field.The optimized cross field not only fits the image boundary features,but also has a certain degree of smoothness.Therefore,the constructed quadrilateral mesh has a better topology.In summary,we mainly focus on the detection and application of geometric features of 3D graphics.For 3D graphics sparse feature detection,we propose a method based on dense coding to extract 3D graphics sparse features,and we propose a three-dimensional tooth model dense coding and feature detection that integrates medical prior knowledge.Besides,the intrinsic geometric features based on the cross field representation data are proposed,and the point cloud upsampling and image quadrilateral meshing are realized under the guidance of the cross field.For different algorithms,we compare the proposed algorithm with the current state-of-the-art methods,verifies the advantages of the method,and realizes multiple application tasks according to its algorithm advantages.