Research On Feature Extraction Technology Of Non-rigid 3D Shape

3D shape retrieval has important theoretical and practical value,aiming to find similar models to a retrieval object in a huge shape database.The key step of the task is to design features for 3D shapes.Non-rigid 3D shapes contain hinge structures resulting in diverse deformations with complex forms,so the search for such shapes is more challenging.There are two main trends in non-rigid 3D shape feature extraction.One is to design hand-crafted features with better discrimination based on exploring shapes' intrinsic properties.The other is to extract features automatically with deep learning while satisfying the lower computational complexity.This paper focuses on designing efficient features and applying them to similarity measures to improve the retrieval accuracy of non-rigid 3D shapes.This paper proposes three features for non-rigid shapes.Firstly,the interior dihedral angle histogram descriptor is designed to analyze the roughness of shape surface,and the retrieval accuracy is greatly improved compared with traditional statistical features.Secondly,the wave kernel signature describes shapes' intrinsic properties using average probability distributions of quantum particles with different energies at given vertices.This paper introduces the eigenvalue normalization method to realize a scale-invariant wave kernel signature,and also improves the retrieval accuracy.Finally,meshes are sampled onto the 2D plane using area preserving sphere parameterization,and then a plurality of spectral features are used to encode points to generate geometric images.3D shape features are obtained by training geometric images through a convolutional neural network,and the retrieval effect based on this feature is improved compared with these coding features.For the above three kinds of features,this paper carries out rich experiments on a number of standard non-rigid 3D shape databases,and compares them with traditional algorithms in multiple angles.Experimental results verify the effectiveness of the proposed algorithms.And a non-rigid 3D shape retrieval system is constructed based on the WebGL to further illustrate the fact.This system can query multiple models similar to the retrieved object in the given databases,and display them in order of similarity.In the end,this paper integrates literature and experiments,and gives some conclusions and prospects,so as to provide reference for researchers in related fields.