Research of Learning and Extracting Manufacturing Quality by Voxel Mapping

通过体素映射学习和提取制造质量的研究

基本信息

批准号：
13650804
负责人：
NAGASAKA Yoshiyuki
金额：
$ 2.37万
依托单位：
Konan University
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
2001
资助国家：
日本
起止时间：
2001 至 2004
项目状态：
已结题

项目摘要

A methodology for extracting and learning geometric features has been studied. At first, an algorithm was developed to create voxels (volume pixels) easily from three-dimensional CAD data. A voxel is a picture element in a three-dimensional coordinate system. Using this algorithm, CAD data such as STL formatted data can be converted to very fine voxels in a few seconds. Then, distance values from the surface of each voxel are calculated. At the same time, a Ds (distance from surface) map can be obtained. In the next step, skeletons of shapes are extracted as polygons from the mapped data.Voxels with the maximum Ds among neighboring voxels are selected as skeleton candidates. After the selection, the candidates are converted to straight lines or circular rings. They are then represented by several vectors and stored as a tree structure. A standard tree involves, for example, four levels and each branch has four descendants. Each parent branch has the same number of descendants. The attributes include such as the scaled volume, connection strength, and scaled X, Y, Z coordinates. These vectors are the input to a skeleton classifier, which is constructed on the basis of the back propagation neural network model. The proposed approach was implemented on a PC machine.A viewer was developed to display the skeletons clearly in three-dimensional shapes. Several parts were selected to demonstrate the classification capability for this methodology. Here, the skeletons extracted directly from the three-dimensional shapes were used. The back propagation neural network model should be trained using some representative shapes. It was found that this method is practically useful through these experiments.

研究了提取和学习几何特征的方法。首先，开发了一种算法来轻松地从三维 CAD 数据创建体素（体积像素）。体素是三维坐标系中的图像元素。使用该算法，CAD 数据（例如 STL 格式的数据）可以在几秒钟内转换为非常精细的体素。然后，计算距每个体素表面的距离值。同时可以获得Ds（距表面的距离）图。在下一步中，从映射数据中将形状的骨架提取为多边形。选择相邻体素中具有最大 Ds 的体素作为候选骨架。选择后，候选对象将转换为直线或圆环。然后它们由多个向量表示并存储为树结构。例如，标准树涉及四个级别，每个分支有四个后代。每个父分支都有相同数量的后代。这些属性包括缩放体积、连接强度和缩放 X、Y、Z 坐标等。这些向量是骨架分类器的输入，该骨架分类器是在反向传播神经网络模型的基础上构建的。所提出的方法在 PC 机上实现。开发了一个查看器来以三维形状清晰地显示骨骼。选择了几个部分来演示该方法的分类能力。这里，使用了直接从三维形状中提取的骨架。应使用一些代表性形状来训练反向传播神经网络模型。通过这些实验发现该方法在实际中是有用的。