Research On3d Model Semantic Annotation Based On Shape Similairty

Semantic annotation is an important part of3D model semantic retrival, the semanticannotation effect directly affect the accuracy and effiency of the retrieval results. At present,3D model semantic annotantion methods can be divided into two categories: one is automaticsemantic annotation, this metnod requires the help of3D model sample set that has beenmarked, and uses statistical learning or machine learning methods to achieve semanticannotation; the other method realizes semantic annotation by relevance feedback-based3Dmodel retrieval, and it not must requies the marked3D model sample set, it realize themapping between3D model low-level features and high-level semantic by relevance feedbacktechnique.Integrate different shape features can complement each other's advantages and describe3D models better, so this paper researched the methods based on mixed shape features, fordata sets with different conditions, to improve the semantic annotation effect and the semanticretrieval accuracy and effiency.In the paper, for the defects of small-scale manual annotation data set and incompletelable information, a method that using graph-based semi-supervised semantic lablepropagation algorithm to expand marked models numnber has been proposed, mentime themethod gives the semantic confidence to the expanded models, and on this basis weightedmeasure learning, then used the distance metrics to achieve multi-semantic annotation. Thismethod used a large number of unlabled samples to participate in the annotation process, andthis effect is better when there is a small amount of manual annotation information.In the paper, for data sets with no label information, an annotation method based oncluster analysis and relevance feedback was proposed. Get the semantic similarities betweenthe3D models using the users' evaluation information for retrieval results in relevancefeedback, and combine it with model features similarities to build a new similarity fusionspace, and achieve the joint clustering based on features and semantic. Meantime useWordNet dictionary to expand users annotation terms, finally achieve multi-semanticannotation of3D model. The method used semantic information in the clustering process,effect has been significantly improved.In the paper, for classified data sets with a small amount of label words, an annotationmethod based on semantic probability model was proposed. The core is to use the probability model which was created by manual annotation lables adjust the similarity measure, toestablish contact between low-level features and high-level semantics of3D models andupadate the probabilistic model dynamically in relevance feedback process, finally improvedthe effect of semantic annotation.