针对反向传播(BP)神经网络用于高分辨率雷达目标距离像分类的问题.讨论了对识别性能产生影响的主要因素:训练算法的选择、输入数据的预处理方法以及神经网络的参数设计。利用4种飞机模型的重点散射源二维分布测试数据和频率步进法得到目标的一维距离像,对于从不同方位角范围内获得的距离像,用BP神经网络对目标识别的性能进行了仿真测试,结果表明选择弹性传播算法或模拟退火弹性传播算法训练网络时具有更好的分类性能,而且对输入样本进行对数变换也有助于提高识别率。

A simulation model of classifying high resolution radar range profiles by BP neural network is presented.The training algorithms,data preprocessing methods and the parameter design of the neural network were main factors which affect the recognition performance.The range profiles are obtained by step-frequency technique using two-dimensional backscatters distribu- tion data of four different scaled aircraft models.Simulations were carried out to evaluate the classification performance with range profiles measured from different aspect angle.The superior performance of the R_PROP(resilient propagation} algorithm and SARPROP(simulated annealing resilient propagation } algorithm over other training algorithms is demonstrated by computer simula- tions.Furthermore,It is shown that logarithm transform is an effective preprocessing method.

TP183

国家部级科研项目