神经网络非均匀性校正算法中初始权值选取方法  

Selection Method of Initial Value of Neural Network for Nonuniformity Correction

作　　者：张科[1] 赵桂芳[1] 崔瑞青[1] 袁祁刚[1] 

机构地区：[1]西北工业大学,陕西西安710072

出　　处：《激光与红外》2007年第3期248-251,共4页Laser & Infrared

基　　金：航天科技创新基金(N4CH008);航空科学基金(04I53067);航空支撑科技基金(05C53005)

摘　　要：针对神经网络对红外焦平面阵列进行非均匀校正时初始权值的选取对收敛速度的影响,提出了三种网络训练时初始权值的选取方法。通过对先验知识总结、黑体训练以及实验标定实现了对网络初始权值的优化。仿真实验表明这三种方法较传统的随机给定法训练收敛速度快、精度高,为神经网络非均匀性校正法走向实时实现提供了重要的理论参考。The selection of the initial value of neural network for nonuniformity correction for IRFPA has great effect to constringency ratio. Aiming at this problem, three methods of selection of the initial value are put forward, which are value given directly, black body trained method and two point method. The initial value of neural network is optimized by these methods. The result of simulation shows that these three kinds of method have much more quick con- stringency ratio and high precision than traditional method, which provides the important theory to real -time realization of neural network for nonuniformity correction.

关 键 词：红外焦平面阵列 非均匀校正 神经网络 初始权值 

分 类 号：TN215[电子电信—物理电子学]