计算机辅助设计与图形学学报
計算機輔助設計與圖形學學報
계산궤보조설계여도형학학보
Journal of Computer-Aided Design & Computer Graphics
2015年
11期
2031-2038
,共8页
施逸飞%熊岳山%谢智歌%徐凯
施逸飛%熊嶽山%謝智歌%徐凱
시일비%웅악산%사지가%서개
几何分析%网格分割%多核学习%超限学习机
幾何分析%網格分割%多覈學習%超限學習機
궤하분석%망격분할%다핵학습%초한학습궤
geometry analysis%mesh segmentation%multiple kernel learning%extreme learning machine
网格分割是三维几何分析的重要问题之一, 它不仅在传统的建模、渲染等方面起到关键作用, 同时也是高层次几何分析的基础性工作.文中提出一种基于多核学习(multiple kernel learning)的快速网格分割算法.多核学习使用多个核函数的组合代替单一核函数, 能够解决网格分割特征多样性和异构性的问题. 给定一组同类别带分割标签的网格模型, 算法首先对网格进行过分割处理, 将三角面片转化为超面片(super-face), 然后使用多核超限学习机训练分割分类器, 最后用该分割分类器对未分割的网格进行分割. 过分割处理能够减少训练样本数量, 进而提高计算效率; 多核学习使分类器能够有效地发现数据间的关系, 使其具有更强大的学习能力. 实验表明, 文中算法不仅计算精度高, 并且能够满足网格分割"实时学习"的计算要求.
網格分割是三維幾何分析的重要問題之一, 它不僅在傳統的建模、渲染等方麵起到關鍵作用, 同時也是高層次幾何分析的基礎性工作.文中提齣一種基于多覈學習(multiple kernel learning)的快速網格分割算法.多覈學習使用多箇覈函數的組閤代替單一覈函數, 能夠解決網格分割特徵多樣性和異構性的問題. 給定一組同類彆帶分割標籤的網格模型, 算法首先對網格進行過分割處理, 將三角麵片轉化為超麵片(super-face), 然後使用多覈超限學習機訓練分割分類器, 最後用該分割分類器對未分割的網格進行分割. 過分割處理能夠減少訓練樣本數量, 進而提高計算效率; 多覈學習使分類器能夠有效地髮現數據間的關繫, 使其具有更彊大的學習能力. 實驗錶明, 文中算法不僅計算精度高, 併且能夠滿足網格分割"實時學習"的計算要求.
망격분할시삼유궤하분석적중요문제지일, 타불부재전통적건모、선염등방면기도관건작용, 동시야시고층차궤하분석적기출성공작.문중제출일충기우다핵학습(multiple kernel learning)적쾌속망격분할산법.다핵학습사용다개핵함수적조합대체단일핵함수, 능구해결망격분할특정다양성화이구성적문제. 급정일조동유별대분할표첨적망격모형, 산법수선대망격진행과분할처리, 장삼각면편전화위초면편(super-face), 연후사용다핵초한학습궤훈련분할분류기, 최후용해분할분류기대미분할적망격진행분할. 과분할처리능구감소훈련양본수량, 진이제고계산효솔; 다핵학습사분류기능구유효지발현수거간적관계, 사기구유경강대적학습능력. 실험표명, 문중산법불부계산정도고, 병차능구만족망격분할"실시학습"적계산요구.
Mesh segmentation is one of the most important problems in 3D geometry analysis. It not only plays an essential role in traditional areas such as modeling and rendering, but also is a foundation of high-level geometry understanding. This paper presents a fast 3D mesh segmentation algorithm based on Multiple Kernel Learning. The Multiple Kernel Learning method deals with the diversity of features in mesh segmentation by using a combination of kernels instead of a specified kernel. Given a set of well-segmented meshes of a cate-gory, the algorithm first over-segments the meshes, then trains a classifier using Multiple Kernel Learning. Af-ter that, the classifier is used to segment other un-segmented meshes. In the algorithm procedure, over seg-mentation reduced the number of training samples, and Multiple Kernel Learning enhanced the ability to reveal the relationships among the data. Results show our algorithm not only outperforms the state-of-the-art mesh segmentation method, but also maintain a high computational speed, which makes it suitable for real-time segmentation learning.