基于非局部均值滤波与神经网络的红外焦平面阵列非均匀性校正算法

摘要

Traditional neural network nonuniformity correction method has the drawback of low convergence speed and is easy to generate ghosting artifacts. To overcome these problems, a neural network nonuniformity correction algorithm based on the non-local means filter is proposed for the infrared focal plane array in this study. To estimate the true image with a higher degree of confidence, the non-local means filter is employed to replace the average filter which is used in the traditional neural network method for its strong ability of edge preservation and global optimization. A variable learning rate is designed in the recursive parameter update process to eliminate the ghosting artifacts more effectively. The performance of the proposed method is tested with two infrared image sequences, which are contaminated with high spatial frequency and low spatial frequency nonuniformity, respectively. Compared with other well-established nonuniformity correction methods, our method has the strength in significantly increasing the convergence speed and meanwhile reducing the ghosting artifacts.%深入剖析传统神经网络非均匀性校正方法收敛速度慢以及易产生“鬼影”现象的主要原因，在此基础上，提出一种基于非局部均值滤波和神经网络的红外焦平面阵列非均匀性校正算法。为了加快收敛速度并减少“鬼影”现象，在神经网络隐含层，利用具有全局寻优且能保持边缘的非局部均值滤波器代替传统的均值滤波器以估计具有更高置信度的真值影像；同时设计可变学习率来自适应地调整每个探测元的非均匀性校正参数的迭代更新过程，以进一步消除“鬼影”。采用两组分别受高空间频率和低空间频率非均匀性干扰的真实红外序列图像进行实验。实验结果表明：相较于目前已有的方法，本文方法不仅具有较快的收敛速度，而且较大程度上抑制了“鬼影”现象的发生。