基于形态学变权神经网络的数据精炼

Data Abstraction Based on Morphological Adjusted-Weight Neural Network

本文研究了模式样本的选取操作要领和典型样本数据的精炼方法 .为改善神经网络普遍存在着误差精度与收敛速率对初始权值颇为敏感甚至过于依赖的缺陷 ,采用感受野的Gabor函数模型来优化网络权值的初始量 .这不仅能为学习过程提供一个良好的开端 ,重要的是将赋予网络模型以具备可塑性优化的基础和适应环境变化的潜力 .算法运用主成份数学分析法对结构元分量的贡献率进行了统计度量与计算 ,以提取出对滤波运算最具影响力的主要分量 ,从而可显著降低数据处理量 ,使运行速度和学习效率更高 .

The selective operation of pattern sample and abstracting method of typical sample data are discussed in this paper. In order to improve obvious shortcoming existing in neural networks in which error precision and convergent rate will be sensitive to initial weight values, even more depending on them, the Gabor function model of visual perception field is applied to optimize initial weight value of neural networks. In this way a good start in learning process can be provided. It is more important to obtain plastic superiority adaptive to complicated alterable environment for the neural network model, and to implement optimal computing principle in which operation load of structuring element weights in morphological filter can be distributed according to their contributive rate.