改进的基于神经网络的非均匀性校正算法

提出了一种结合图像匹配和神经网络算法的焦平面阵列非均匀性校正算法。算法首先用最新的校正系数对图像进行非均匀性校正，输出校正结果；然后对相邻两帧图像进行匹配，估计出相邻帧之间图像的运动量；最后用神经网络算法分别对校正系数进行正向和反向自适应更新。采用图像匹配技术保证了校正系数更新时不会引起场景的模糊，采用校正系数双向更新策略可以保证每帧都能对每个像元的系数至少进行一次更新，与常用的神经网络校正算法相比，降低了对场景统计特性的要求，收敛速度较快。使用模拟添加噪声和采集的红外图像序列对算法进行仿真验证，结果表明，给出的算法校正效果优于常用的神经网络非均匀性校正算法。

Improved nonuniformity correction algorithm based on neural network

An improved nonuniformity correction (NUC) algorithm combining image matching and neural network(NN) for infrared focal plane array sensors was presented. Firstly, nonuniformity of the FPA response was removed by NUC compensation. Then, motion parameters of the image were estimated by matching pairs of image frames. Finally, coefficients were adaptively updated according to bidirectional-renew strategy based on neural network. Image matching technique could effectively avoid faintness when coefficients were updating. Additionally, the bidirectional-renew strategy was used to guarantee coefficients of each pixel be calculated at least once when new image frame came. The new algorithm used image matching technique to get scene motion information, and used neural network for coefficients bidirectional-renew strategy. It had a lower statistical overhead on scenes and approached convergence more quickly than the often used neural network based NUC algorithms. A theoretical analysis was performed on a collection of infrared image frames to study the accuracy of the new NUC algorithm. It proves that it has higher-quality correction ability than simple neural network based NUC algorithm.