| 用于图像目标识别的神经网络方法 |
| |
| 引用本文： | 沈定刚,戚飞虎.用于图像目标识别的神经网络方法[J].系统工程与电子技术,1994(9). |
| |
| 作者姓名： | 沈定刚 戚飞虎 |
| |
| 作者单位： | 上海交通大学光纤所 |
| |
| 基金项目： | 国家攀登计划认知科学（神经网络）重大关键项目 |
| |
| 摘 要： | 本文提出了一种全局最优的神经网络（ＦｕｌｌＤｏｍａｉｎＯｐｔｉｍｕｍＮｅｕｒａｌＮｅｔｗｏｒｋ）模型用于目标识别。通常所设计的神经网络不能保证全局最优，使得网络不一定收敛到期望样本点上。本文的模型采用了先设计稳定点、再构造吸引域的方法，提高了网络的识别正确率及速度。针对图像识别中矢量维数大的实际，提出了一种不变性方法，使得样本维数下降而分类距离保持不变。同时又证明了网络的收敛性、收敛速度及映射保距等。计算机模拟结果表明，网络对噪声或缺损图均能正确识别。
|
| 关 键 词： | ￣＋神经网络，模型，目标识别，图像识别，模拟。 |
A FDO Neural Network Model and Its Application to Image Recognition |
| |
| Shen Dinggang and Qi Feihu.A FDO Neural Network Model and Its Application to Image Recognition[J].System Engineering and Electronics,1994(9). |
| |
| Authors: | Shen Dinggang and Qi Feihu |
| |
| Abstract: | A foll domain optimum neural network(FDONN)and its application to image rec-ognition are proposed in this paper.A neural network generally cannot converge to a global mini-mum.In this paper,a method of devising stable points firstly and basins of attraction laterly is used,by which the speed and correctness of recognition are improved.Owing to the fact that the dimen-sions of image are large,an invariant transformation method is presented to decrease the dimen-sions of image without changing the distance between them. The properties of the neural network,such as quality and speed of convergence and invariance of mapping, are proved,too. Several com-puter simulation examples are given. The recognised images include noise-added images and defec-tive images. |
| |
| Keywords: | Neural network Model Target recognition Image recognition Simulation |
| 本文献已被 CNKI 等数据库收录！ |
