权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Intelligent image processing based on distributed-cooperative environment

基于分布式协作环境的智能图像处理

基本信息

批准号：
14580435
负责人：
WATANABE Eiji
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：
2002
资助国家：
日本
起止时间：
2002 至 2003
项目状态：
已结题

项目摘要

(1)Intelligent learning of neural networks for image processing :(a)An image segmentation method using multiple multi-layered neural networks has been newly introduced and an image segmentation method is proposed based on the differences among the compression abilities in multiple NNs. The segmentation results by the proposed method have been concretely shown by some experiments.(b)A cooperative learning algorithm for multi-layered neural networks is proposed.The proposed algorithm has the following procedures ; (i)parallel learning of each NN with different initial weights by using the back propagation learning algorithm, (ii)exchange of weights so as to minimize the unbalance among the group of NNs and improve the learning performance of them. We have applied the proposed algorithm for a classification problem of Japanese-Kanji fonts and have shown its behavior and effectiveness.(2)Image Restoration :(a)Considering that a partial image which contains some edges tends to become a bina … More ry image, we introduced a new evaluation function that takes a maximum for binary images and proposed an automatic extracting method of an edge area from a degraded image using the function.(b)We discussed the problem to estimate an original image and a point spread function from only a degraded image. We showed that there exist finite solutions if a degraded image has a compact support. Then, we proposed an algorithm to find all solutions.(3)Considerations on the structure of a distributed and cooperative computational environment :Our distributed and cooperative computational environment is based on a PC cluster, Unix-like operating systems, and MPI. We proposed a method of the dynamic load balancing on such environment. In the method, we regarded a cluster and a PC cluster system as an autonomous agent and a multi-agent system respectively. The cluster is designed using the meta-level architecture which consists of base level for usual computations and meta level for global computations. According to the method, the dynamic load balancing on such environment is performed as the communication and load balancing on the multi-agent system. Less

（1）用于图像处理的神经网络的智能学习：（a）已经新引入了使用多个多层神经网络的图像分割方法，并且提出了基于多个神经网络中的压缩能力之间的差异的图像分割方法。实验结果表明，该方法的分割效果较好. (b)A提出了一种多层神经网络的协同学习算法，该算法利用反向传播学习算法对具有不同初始权值的各个神经网络进行并行学习，并通过权值交换来减小神经网络组间的不平衡性，从而提高神经网络的学习性能.我们已经应用所提出的算法的分类问题的日文汉字字体，并已显示其行为和有效性。（2）图像恢复：（a）考虑到包含一些边缘的部分图像倾向于成为二值图像， ...更多信息针对退化图像，引入了一种新的二值图像取极大值评价函数，并提出了一种利用该函数自动提取退化图像边缘区域的方法。(b)We讨论了仅从退化图像估计原始图像和点扩散函数的问题。我们发现，存在有限的解决方案，如果退化的图像有一个紧凑的支持。然后，我们提出了一个算法来找到所有的解决方案。（3）分布式协同计算环境的结构考虑：我们的分布式协同计算环境是基于PC集群、类Unix操作系统和MPI的。提出了一种在这种环境下实现动态负载平衡的方法。在该方法中，我们把一个集群和一个PC集群系统分别看作一个自治Agent和一个多Agent系统。该集群采用元级体系结构，由基本级和元级组成，基本级用于常规计算，Meta级用于全局计算。根据该方法，在这样的环境中执行动态负载平衡，作为多代理系统上的通信和负载平衡。少