摘要
微晶玻璃颜色分类是最终控制产品质量的重要步骤 .作者改进了传统ART2网络的学习算法 ,借用典型向量的概念 ,以模式的近似均值作为典型向量来快速学习新模式 .改进学习算法极大地改善了ART2网络的模式漂移现象 ,而且能缩短搜索振荡过程 .文中分析了微晶玻璃颜色分量的统计信息 ,经过适当变换将高维颜色特征映射到 16维特征空间中的一个超平面上 .以超平面上的特征点作为改进算法ART2网络的输入进入网络分类器进行学习分类 .实验证明改进算法网络用于微晶玻璃颜色分类时 ,运行正确、可靠 。
Color grading is one of key stages for the final quality control of glass_ceramic products. In this paper, the learning algorithm of the classical ART2 has been modified. The concept of typical vector is used to modify the learning algorithm, that is, rapidly learning the newest pattern by taking approximate mean vector as typical vector. The modified learning algorithm can greatly improve the pattern_shifting problem of classical ART2, and is able to shorten the searching process. The statistic distribution of samples' color elements H, S, V is analyzed and they are mapped to a super_surface of 16 dimensions via proper transformation, thus decreasing the feature space dimension. The mapping points of samples in super_surface have been sent to ART2 classifier as input data of net learning. The experimental results show that ART2 classifier based on modified learning algorithm is effective, reliable and has fairly high recognition correctness when it is used to glass_ceramic color grading.
出处
《华南理工大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2003年第1期74-78,共5页
Journal of South China University of Technology(Natural Science Edition)
基金
广东省科技攻关项目 (2KM 0 0 6 0 8G)
