基于BP神经网络的大视场成像畸变校正方法

BP NN-based method for lens distortion correction of large-field imaging

提出了基于BP神经网络的畸变校正方法,实现了在不确知畸变数学模型情况下的高精度校正。应用平行光管法测得900组光点坐标数据作为BP神经网络的训练样本,通过训练得到一个隐含层有6个节点的3层BP神经网络模型,然后使用77组数据检测了该神经网络的校正效果,并与使用相同检测数据的双线性插值方法进行了比较。实验表明,神经网络较插值方法精度高,误差数据分布的规律性好,而且与插值方法相比,它避免了误判区域的问题,便于校正。

A method based on BP neural network (NN) is proposed in this paper, which realizes high-precision correction without knowing the mathematic model. 900 groups of light-point data are obtained as training samples by parallel light pipe method. A three-layer NN model with 6 nodes in the hidden layer is achieved and its correction effect is examined by 77 groups of data. The dual linear interpolation method is applied to the same 77 groups of data in comparison. The conclusion is that comparing to interpolation method, the BP NN method has higher precision and the BP NN method can avoid the problem of error estimation region in the interpolation method.