Abstract: A learning algorithm for two-weight neural networks is introduced which based on error back propagation. It can synchronously determine the direction weight, core weight, index parameter, and learning factor parameters. Various hypersurfaces feature can be simulated by properly tuning these parameters. In the training of two-weight network, this paper compares the method with BP algorithm through the segment of portraits. At last, the writer applies the two-weight neural networks to segment the vehicle license plate, and achieves good image segment effect.

Key words: BP algorithm, two-weight neural networks, image segmentation

摘要： 采用基于误差反向传播的双权值神经网络学习算法，同时确定核心权值、方向权值以及幂参数、学习率等参数，通过适当地调节这些参数，从而实现尽可能多种不同超曲面的特性。在对双权值网络进行训练时，通过对人物头像的分割，将该算法与带动量项BP算法进行了比较。最后将双权值神经网络成功地运用于车牌号码等图像的分割工作中，取得了良好的图像分割效果。

关键词: BP算法, 双权值神经网络, 图像分割