摘要
提出了一种立体神经视觉系统中零件识别的学习方法 ,与标准的BP算法对比有两点改进 :①用变尺度方向代替负梯度方向作为搜索方向 ;②用可变的最优学习率来代替不变的学习率 .采用上述 2个改进后 ,训练速度和收敛性都有较大的改善 .实际应用表明 ,所提出的学习方法的训练速度。
A learning mechanism for parts recognition(LMPR) in a stereoscopic nuro vision system is presented. It differs from the mechanism used in the standard back propagation (SBP) neural network in two ways. First, the searching direction is changed from the negative gradient direction to the variable metric direction. Secondly, the constant learning rate is changed to a variable optimal learning rate. With the combination of variable metric direction and variable optimal learning rate, the speed of the training process is greatly improved and the convergence is assured. Application examples are presented. The results have indicated that the proposed LMPR is superior in comparison with the SBP in the areas of learning speed, convergence and stability.
出处
《中山大学学报（自然科学版）》
CAS
CSCD
北大核心
2000年第3期25-29,共5页
Acta Scientiarum Naturalium Universitatis Sunyatseni
基金
广东省博士后基金资助项目
关键词
立体神经视觉
学习方法
神经网络
零件识别
装配
stereoscopic neuro vision
learning mechanism
neural networks
variable metric direction
variable optimal learning rate
