Classification And Retrieval Of Non-rigid 3D Models Based On Heat Kernel And Deep Learning

With the rapid development of multimedia technology and computer hardware and software,a huge number of 3D shapes are widely used in many fields such as design reuse,intelligent production.Therefore,it is an urgent need in every field to solve the difficult problems in the classification and retrieval of 3D shapes and to realize effective management of massive 3D shapes.The effective management of 3D shapes can improve the efficiency of design and production,improve the reuse rate of the original model,reduce the time and labor cost.In recent years,Artificial Intelligence(AI)has achieved rapid development both in theory and in application.In particular,the theoretical method of Deep Learning(DL)has been introduced to solve many difficult problems on AI,and the application of AI to various fields is realized.In the research field of 3D shapes classification and retrieval,improving the accuracy,applicability and intelligentization of the algorithm is an important direction of research and development.In the 3D shapes classification and retrieval technology,we need to solve the research problem of feature extraction,classification and retrieval of 3d model.The purpose of this study is to efficiently find the required model in mass data.At present,most of the research work on the classification of 3D shapes is aimed at rigid body model,and its algorithm is difficult to apply to the non-rigid 3D shapes with variable attitude.This paper is based on the heat kernel which describes the thermal diffusion law,study the feature extraction,classification and retrieval methods for the non-rigid 3D shapes,a feature extraction algorithm of NSIHKS(New Scale-invariant Heat Kernel Signature)is presented.In combination with the deep learning method,a classification and retrieval framework of a non-rigid 3D shape is constructed.The research results of this paper are of great significance for product design reuse,e-commerce and other fields.The main research contents are as follows:(1)Summarized the existing classification and retrieval methods of general 3D shapes,rigid 3D shapes and non-rigid 3D shapes.The NSIHKS feature extraction algorithm is proposed,and the advantages of this feature in the application of non-rigid3 D shapes are analyzed.As the local feature of the vertex of the shapes,HKS(Heat Kernel Signature)is equidistant and constant volume invariance.The application is effective in the classification and retrieval of non-rigid 3D shapes,but its shortcoming is easily affected by the scale of the shapes.In this paper,the discrete Fourier transform is used and draw on the experience of extracting the feature SIHKS from non rigid shapes.Time scale and feature dimension are optimized.An NSIHKS feature extraction algorithm for non-rigid 3D shapes is proposed.(2)According to the k-means algorithm,the suitable clustering center for shape dataset is generated by modifying and calculating.The NSIHKS feature is transformed into a new descriptor for the shape.A measure index based on similarity of Hausdorff distance model is presented for classification and retrieval of non-rigid 3D shapes.(3)Based on DL theory,the convolution neural network(CNN)is designed to extract the depth feature of non-rigid 3D shapes.A classification and retrieval framework for non-rigid 3D shapes is proposed to further improve the accuracy of classification and retrieval.Experimental results show that this method has good performance in non-rigid3 D shapes classification and retrieval.