在红外成像制导应用中, 为满足长周期免拆卸贮存的应用需求, 红外导引头非均匀性的研究越来越多的集中于采用自适应的校正方法来代替传统的参考源的非均匀性校正方法。针对传统基于神经网络的自适应非均匀性校正算法容易造成“鬼影”的问题, 提出了一种改进的红外导引头成像自适应非均匀性校正算法。该方法在传统神经网络非均匀性校正的基础上, 进行了 4点实用化的改进: 首先, 通过对图像运动判断, 避免场景静止时的过学习; 其次, 采用自适应学习率, 避免细节丰富区域的过学习; 然后, 利用双边滤波求期望目标的评估, 减少细节的损失; 最后, 通过判断误差函数的波动量来决定是否对偏置进行更新。实验结果表明, 该方法在校正精度、收敛速度和稳定性方面均优于传统的神经网络校正算法。

In the area of nonuniformity correction of infrared image guidance, to satisfy the demand of long period no-take-down reserve, more and more research is focusing on adaptive nonuniformity correction instead of reference nonuniformity correction. Aiming at the problem of traditional adaptive nonuniformity correction easy to produce artificial ghost, an improved adaptive nonuniformity correction method is proposed. The scene-based adaptive nonuniformity correction algorithm is based on traditional neural network nonuniformity correction, but has four applied improvements. Firstly, it avoids the over learning when the scene is still by the movement distinguishing of images. Secondly, it avoids the over learning of detailed area by adopting adaptive learning rate. Thirdly, it reduces the loss of details by using bilateral filter to get the estimation of the expectation. Finally, it updates the offset according to the variation of the error function. Experiments show that the proposed algorithm is better than traditional neural network nonuniformity correction in correction precision, convergence speed and stability.