基于TOP-HAT滤波器算子的红外弱小目标检测算法

Optimized Design of Improved TOP-HAT Filter Based on Genetic Algorithms of Neural Network

针对红外序列图像中运动弱小点目标的检测问题,设计了一种基于改进神经网络优化的修正T op-H at形态学滤波器算子。其中形态学滤波器的结构元素采用两层前馈神经网络,通过大量样本训练优化,将T op-H at运算作为一个整体当作一层,输出层节点定义为T op-H at运算后图像矩阵的最大值。实测数据的处理结果表明:针对低信噪比(RSN≈2)图像,在虚警概率≤5%情况下,优化的修正T op-H at形态学滤波器算子对复杂图像检测概率≥75%,与固定结构元素的T op-H at形态学滤波器相比检测概率提高了近8%,算法的运算时间仅增加了0.7m s。

An improved morphological Top-Hat filtering operator is designed based on the upgraded neural network for detecting the moving spot target in infrared image sequences. Two-layer feedforward neural network is adopted in structural element of the morphological Jilter. Through training and optimizing of a large quantity of samples, Top-Hat operator is taken as a whole layer, and the output layer nod is defined as maximum value of the image matrix after Top-Hat operation. The operator optimized with two-layer neural network can successfully suppress both background and noise. The output is the value of the structural element. Experimental results of the actual measurement show that by using the improved Top-Hat operator, the detection probability of images with low RsN（R SN≈2） is higher than 75% when false alarm probability is lower than 5%. Compared with fixed Top-Hat filter, the detection probability is improved by nearly 8%, and the time for calculation is only increased by 0.7 ms.