基于卷积神经网络的合成孔径雷达图像目标识别

为了解决现有合成孔径雷达(SAR)图像目标识别算法识别率不高、泛化能力不足的问题,提出一种基于卷积神经网络的SAR图像目标识别模型CMNet网络。通过设计针对SAR图像特点的特征提取网络,在损失函数中引入中心损失与Softmax损失联合监督训练过程,兼顾类内聚合和类间分离,提高算法精度和泛化能力。网络模型中所有卷积层后引入批量归一化层加快模型收敛速度、防止过拟合。实验使用美国运动和静止目标获取与识别数据库进行测试,10类目标平均识别率达到99. 30%。结果表明,提出的CMNet网络模型具有较高的识别率和泛化能力,在公开数据集上取得较好结果。

Target Recognition Using Convolution Neural Network for SAR Images

In order to solve the problem that the existing SAR images target recognition algorithm has low recognition accuracy and low generalization ability, this paper proposes a framework named CMNet based on convolutional neural network for SAR images target recognition model. By designing the feature extraction network for the characteristics of SAR images, introducing center loss in the loss function to construct joint loss with softmax loss for supervision training process, the intra-class aggregation and inter-class separation are considered to improve the accuracy and generalization ability of the algorithm. Batch normalization is introduced after all convolutional layers in CMNet to speed up the convergence of the model and prevent overfitting. The experiments was conducted on Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark data set. The results show the average recognition rate of the 10 categories of targets reached 99.30%. It is concluded that the proposed CMNet model performs high recognition accuracy and generalization ability and gets promising results on public data sets.