一种基于实例学习的三维模型检索匹配方法

3D shape retrieval method based on case learning

传统的基于内容的三维模型检索的相似性度量方法主要借鉴二维图像检索中所采用的距离度量算法,达到比较两三维模型相似度的目的,该做法限制了模型间匹配的广度.针对这种单核匹配的限制,提出了一种新的多核匹配方法.利用图论中两点间的最短距离的思想,得出两模型最相似那他们的距离最短,因此查询样本跟匹配的样本存在的通路上的模型节点能影响他们的相似度,这样就提高了匹配的广度和精度.同时在已有的特征提取基础上,把标签繁衍算法应用到最短距离求解中,并将基于实例学习的K近邻方法引入到模型匹配算法中,实现了半监督学习,提高了系统的查准率.

Most traditional similarity matching methods of content-based 3D shape retrieval mainly has drawn distance metric used in two-dimensional image retrieval algorithms in order to measure similarity between two 3D models. But this approach limits the matching breadth of models. This paper proposes a multicore similarity matching method to enlarge the matching scope. Based on the idea of the shortest distance between two points in Graph theory, two of the most similar 3D models will have shortest distance. Therefore, the node on the path between the query shape and the matching shape can affect their similarity. So the new method can improve the matching breadth and precision. Meanwhile, based on the existed feature extraction methods, the label propagation algorithm is used into the solution of shortest distance and the k nearest-neighbor method based on case learning can also be introduced into similarity matching algorithm. It can realize the semi-supervised learning algorithm and improve the precision of algorithm.