改进的BP算法在多目标识别中的应用  

Multi-target recognition with improved BP algorithm

作　　者：许廷发[1] 张敏[2] 顾海军[1] 宋建中[1] 韩广良[1] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130022 [2]长春师范学院数理学院,吉林长春130024

出　　处：《光学精密工程》2003年第5期513-515,共3页Optics and Precision Engineering

基　　金：中国科学院知识创新工程青年基金资助项目(No.Q01H01)

摘　　要：为了实现对多目标的识别,提出了一种改进的BP算法。采用在步长函数中加入变动量因子,即当连续两次迭代其梯度方向相同时,步长加倍;当连续两次迭代其梯度方向相反时,步长减半。由于步长在迭代过程中自适应进行调整,使误差函数E在超曲面上的不同方向按照步长向极小点逼近,实现了对目标函数的优化。给出了在目标函数最优时的BP网络对三种飞机测试集的识别结果,其收敛速度比传统BP算法快4倍以上,表明该方法能够有效地用于多目标的识别。An improved BP algorithm has been proposed for realization of multi-target recognition. The principle followed is addition of an alternating momentum factor to the step function, i.e., the step length is doubled when the gradient directions of two consecutive iterations are the same, and the step length is halved when the gradient directions of two consecutive iterations are opposite. Due to the adaptive adjustment of step length in the iteration process, error function E approaches the minimum point in different directions on the super curved surface in accordance with the step length to realize the optimization of goal function. The recognition rates achieved using the improved BP algorithm with optimized goal function are better than 98%, and the improved BP algorithm's convergence speed is four-times fast than the convergence speed of traditional BP algorithm.

关 键 词：神经网络 目标识别 BP算法 人工视觉 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]