一种顾及粗差的径向神经网络高程曲面拟合法

Gross Error Considered Height Surface Fitting Using Radial Neural Network

针对径向神经网络高程曲面拟合对训练样本点的粗差较为敏感的特点,从数字影像处理的原理出发,将径向神经网络拟合中含粗差训练样本点得到的格网点高程映射到灰度影像,利用影像高斯差分算法探测粗差,并采用最小二乘移动曲面法修正原始高程训练样本点中的粗差,然后再次对修正后的样本点进行径向神经网络拟合,从而实现可抵抗粗差的径向神经网络高程曲面拟合。以测量机器人采集的高程数据为例进行了实验,达到了较好的拟合结果和精度,验证了方法的鲁棒性与有效性。

Height surface fitting by radial neural network is sensitive to gross error of training samples,aapproach of gross error considered height surface fitting using radial neural network is introduced based on techniques of digital image processing,all height grid points by radial neural network with gross error are mapped to gray image,image difference of Gaussian and least square moving curved surface algorithms are implemented for gross error detection and correction,and implementing height fitting by radial neural network with gross error corrected discrete height points.To verify its feasibility,we implemented height fitting by robust radial neural network with the data from surveying robot.The results show that the method can achieve better accuracy,has strong robustness and feasibility.