摘要
非制冷红外焦平面的非均匀性对红外系统的图像质量造成严重影响。神经网络的自适应调节性优于传统的定标校正方法,成为研究热点。但是传统的神经网络存在期望值不准确、误差函数精度不高和学习速度不适应网络变化的缺点。本文将目标像元与其4邻近像元的像素值进行比较,按偏差值的大小进行排序,再增加权系数来计算期望值;文章又分析了神经网络出现的局部极小问题,在原有的误差函数基础上引入了隐层饱和度的计算式;并提出了根据总误差值之比来调节学习速度。经仿真实验表明,新算法较好地降低了非均匀度。
The non-uniformity of URPA has serious effects on image qualities of infrared system.The adaptive control ability of neural network,which is superior to traditional calibration algorithm is becoming a focus.However,the traditional way has some disadvantages,such as the inaccuracy of expected value,the short precision of error function and the learning rate which is not adapted to the changing of network.In this paper,we compare the target pixel with other four pixels arounding the target one,sort them according to the values of differences,then count expected values by adding weight coefficients.In the article,we also analysis the problem of extreme minimal in some parts in the neural network,introduce a counting function of hidden layer saturation on the basis of the original error function,and propose that moderating the learning rate according to ratio of error values.The simulating experiment indicates that the new algorithm fairly reduces the non-uniformity.
出处
《激光与红外》
CAS
CSCD
北大核心
2010年第10期1111-1115,共5页
Laser & Infrared
基金
河北省科学技术研究与发展计划项目(No.05213503D)资助
关键词
非均匀性
神经网络
误差比较
自适应调节
non-uniformity
neural network
errors comparison
adaptive moderate ability
