基于缓变场景的复杂神经网络非均匀性校正

针对遥感数据定量化应用对多元红外探测器非均匀性校正的高精度需求，提出了一种基于缓变场景的复杂神经网络非均匀性校正算法，在两点校正基础上，进一步降低图像非均匀性。与经典BP神经网络非均匀性校正及其改进算法相比，复杂神经网络非均匀性校正算法突破单一层学习神经元结构限制，采用双层学习神经元结构，第一层学习神经元采用大邻域中值滤波作为期望函数，第二层学习神经元采用小邻域均值滤波作为期望函数，通过多层学习神经元配合，兼顾非均匀性校正效果并避免图像边缘模糊。经实拍红外图像非均匀性校正实验证明，复杂神经网络校正与经典BP神经网络校正相比，取得了更好的非均匀性校正效果，非均匀性评价结果UN=0.75%，次优算法非均匀性评价结果为UN=0.77%，与次优算法相比，复杂神经网络校正算法有更高的像素平均梯度评价值，达到28.49，而次优算法只有28.30，非均匀性评价结果低并且像素平均梯度高说明复杂神经网络校正算法在降低图像非均匀性同时更好地避免了图像边缘模糊：这是复杂神经网络校正算法采用双层学习神经元的特点与优势。方法改善了多元红外探测器非均匀性校正的效果，提高了红外遥感图像的质量，促进了红外遥感的应用。

NON-UNIFORMITY CORRECTION BASED ON COMPLEX NEURAL NETWORK BASED ON FIXED SCENE

Non-Uniformity is a serious problem for the infrared focal plane array,for reducing the non-uniformity of the infrared focal plane,a new Non-Uniformity correction algorithm is suggested,it’s called non-uniformity correc?tion algorithm based on complex neural network based on fixed scene. It’s different with the Non-Uniformity correc?tion algorithms based on traditional BP neural network,which has only one learning-layer,the new algorithm uses two learning-layers,that has many advantages. The traditional Non-Uniformity correction algorithm based on tradi?tional BP neural network blurs the image more when the more better correction result is the goal,however,the Non-Uniformity correction algorithm based on complex neural network uses two learning-layers,in the first learning-layer,the medial filter with a large scale neighbor data is used to generate the theory output data,and in the second learning-layer,the average filter with a small scale neighbor data is used to generate the theory output data,by using two learning-layers,not only the non-uniformity reduces,but also the image edge is kept. In the experiment,the Non-Uniformity correction algorithm based on complex neural network has the best result,the UN=0.75% for cor?rected image,and the UN=0.77%for the second best image,however,the image with UN=0.75%has a bigger aver?age grads value than the image with UN=0.77%,that means the Non-Uniformity correction algorithm based on com?plex neural network has the advantage both on reducing the Non-Uniformity and avoiding the image edge bluring.