基于多特征的BP神经网络LSS目标识别方法

A BP Neural Network Recognition Method Based on Multi Features for the LSS Target

为更好地从可见光探测图像中对"低慢小"无人机目标进行识别和分类,对含有旋翼无人机、固定翼无人机和城市飞鸟这三类目标的图像进行灰度化、二值化等预处理,建立了目标圆形度、Hu不变矩、仿射不变矩特征提取模型以及多特征融合模型。在此基础上,提出了一种基于多特征的BP神经网络目标识别方法,构建包含输入层、隐含层、输出层的三层神经网络训练模型,并明确了该模型的训练过程。选取了三类目标图像,以其中每类各150张作为样本训练集,每类各50张作为样本测试集,通过计算损失函数来判定目标类别。结果表明:上述方法的识别率达92.67%,可实现对城市空域环境的"低慢小"目标的识别。

In order to better recognize and classify the low, slow and small(LSS) UAV targets from the visible-light detection images, gray-scale and binary preprocessing were carried out for the images containing three kinds of targets, i.e. rotorcraft UAV,fixed-wing UAV and urban birds. The feature extraction models of target roundness, Hu moment invariant, affine moment invariant and multi-feature fusion were established. On this basis, a Back Propagation(BP) neural network target recognition method based on multi-feature was proposed, and a three-layer training model of neural network including input layer, hidden layer and output layer was constructed, and the training process of the model was clarified. Three types of target images were selected, 150 images of each type were used as training set and 50 images of each type as testing set. The target category was determined by calculating loss function. The results show that the recognition rate of this method is over 92%,and it can realize the monitoring of the LSS target at of urban airspace environment.