Research On 3D Model Semantic Annotation Based On Content Retrieval

As the fourth multimedia data after sound,image and video,the 3D model has a wide spread and application space in the Internet.However,due to the existence of the semantic gap,the 3D model has the problem that high-level semantics and underlying features are difficult to convert.Funded by the National Natural Science Foundation of the State(No.:61502094)and the Natural Science Foundation of Heilongjiang Province(No.: F2016002).In order to solve this problem and improve the recognition and retrieval efficiency of 3D models,this paper aims at the rapid identification and accurate annotation of 3D models.Based on the comprehensive comparison of the existing 3D model semantic annotation methods,explore the research of semantic annotation method is carried out.The specific research contents are as follows:(1)Aiming at the problem of automatic semantic annotation of three-dimensional model.An Annotation Method of 3D Model Based on Neighbor Voting(AMNV)is proposed.This method improves the traditional k-nearest neighbor method and applies it to the batch automatic labeling method of 3D model.By extracting the D2 distribution of the 3D model as the external shape feature of the 3D model,and use the EMD method to calculate the similarity distance between the label target and the data set,and the k neighbor feature models with similar features are used.The included known tags are used to vote for the labeling results of the labeling model,and finally the automatic semantic annotation of the 3D model is realized.(2)Aiming at the problem that the accuracy of existing annotation results is not high and too single,this paper proposes a Diverse 3D Model Annotation(DMA).On the basis of the existing calculation method of 3D model semantic annotation,the diversity results of the 3D model are optimized.The method proposed in this paper firstly defines the diversity measure of the 3D model,and combines the semantic similarity calculation between the 3D model label and the label.By comprehensively considering the diversity and accuracy of the label,the labeling result below the label correlation expectation will be used.Reorder to maximize the expected label accuracy.New labeling results are generated by using new label sorting to realize the diversity optimization of 3D model semantic labeling.(3)This paper introduces a three dimensional model semantic annotation optimization method based on manual query related feedback.The method mainly relies on the user's further optimization of the query result of the three dimensional model.The user actively queries the relevant annotation results and performs binary confidence scores on the results,so as to optimize the correlation between the existing annotation results and the true semantics of the 3D model,and finally improve the annotation results of the 3D model.In order to meet the requirements of existing 3D model retrieval and use,this paperdesigns and develops a 3D model semantic annotation system based on 3D model semantic annotation and retrieval.Based on the management and use of 3D models,the system complements the content semantic recognition and shape feature retrieval functions of 3D models,which enables users to be more efficient and convenient when searching and using3 D models.