稳像中基于BP神经网络的颤振预测及改进

Vibration forecasting using BP neural network for image stabilization and an improving method

为了解决航拍过程中的图像抖动问题,研究了机载相机的颤振规律;提出在稳像过程中,利用BP(BackPropagation)神经网络的函数逼近功能对相机颤振规律进行模拟,预测相机颤振矢量的方法;针对单个BP神经网络稳定性较差且精度较低的问题,提出在预测网络上增加一个误差校正网络以提高预测精度的方法.该方法使用误差校正网络对预测网络输出的结果进行二次预测、补偿,提高了网络系统的稳定性和计算精度.仿真实验表明:在训练样本相同的情况下,预测网络和误差校正网络相结合的方法能够对相机颤振矢量进行高精度预测,且运算速度较快,满足了机载相机实时稳像的需求.

The vibration characteristic of airborne camera was studied to solve the image vibration in aerial photography.A method based on the ability of function approximation of BP neural network to simulate the vibration characteristic of airborne camera and predict the vibration displacement vectors during image stabilization was proposed.The inherent defects of BP neural network are its poor stability and low precision which cannot be accepted in the application of image stabilization.To overcome these problems,a new method combining two networks named prediction network and error correction network was proposed.The later network performs further prediction and compensation on the outputs of the former one,and thus optimizes the property of the network system.Experimental results show that with the same training samples,the combined network system is more stable and the outputs are of higher precision than that of a single network,and the computing is also fast,all of which meet the demands of real-time image stabilization in aerial photography.