Study on Automated Constuction of Three Dimensional Image Processing Procedures

本项目的目的是开发一种基于用户给出的称为样本图形的3D二值模式的3D图像处理程序的自动构建方法。在项目的前半部分，收集和评估了3D图像处理的基本算法和基本程序，为我们的方法提供了知识库。在后半部分，以上述知识库为基础，提出了一种自动构建三维图像处理程序的方法。分别对海量图案、表面图案和线性图案提取程序进行了实施，并使用人工和实际3D图像进行了评估。构建的基本流程包括以下三个步骤：(1)程序框架的选择；(2)基于程序框架的详细程序构建；(3)详细程序的绩效评估。特别地，在第二步中，在程序框架中的某些合理点处自动生成被称为预期图像的中间目标，然后通过评估处理结果与对应的中间目标之间的相似度来顺序地确定详细的程序。对人工三维图像和实际三维图像的实验表明，该方法可以自动构建一个从三维灰度图像中提取与给定样本图形相似的图案的过程，基于以上结果，可以开发出三维图像处理专家系统。要做到这一点，应该研究从“多个样本数字”和/或“负样本数字”构建程序。

The purpose of this project is to develop a method for automated construction of a three-dimensional (3D) image processing procedure based on a 3D binary pattern called "sample figure" given by a user as a goal of processing.In the first half of the project term, collection and evaluation of basic algorithms and basic procedures for 3D image processing were performed for providing a knowledge database for our method. In the second half, based on the above knowledge database, a method for automated construction of 3D image processing procedures was developed. Implementation was performed individually for massive pattern, surface pattern and linear pattern extraction procedures, with evaluation by using both artificial and practical 3D images. The basic flow of construction consists of following three steps ; (1) selection of "procedure frame", (2) construction of detail procedure based on procedure frame and (3) performance evaluation of detail procedure. Especially, in the second step, intermediate goals called "expected images" are automatically created at some reasonable points in a procedure frame, and then a detail procedure is fixed sequentially with evaluating similarity between a processing result and the corresponding intermediate goal. In experiments with both artificial and pracitical 3D images, it was shown that the method can automatically construct a procedure to extract a similar pattern to the given samaple figure from the 3D gray image.In future, based on the above results, development of an 3D image processing expert system can beexpected. To do this, procedure construction from "many sample figures" and/or "negative smaple figures" should be investigated.