深度卷积神经网络在迁移学习模式下的SAR目标识别

Target recognition using the transfer learning-based deep convolutional neural networks for SAR images

合成孔径雷达(synthetic aperture radar,SAR)自动目标识别过程主要包括目标特征提取和分类器训练两个步骤。提出一种基于深度卷积神经网络(deep convolutional neural networks,DNNs)的SAR自动目标识别方法,使用一类优化的DNNs网络结构对SAR图像目标进行分类训练。该网络结构自动提取目标类别特征,避免人工预选取特征方法带来的不标准性。在DNNs网络模型训练过程中引入迁移学习的概念,以防止结果陷入局部最优解和加快模型参数的训练。最后使用美国运动和静止目标获取与识别MSTAR数据集进行试验,给出该方法与其他分类方法结果的对比,证明其取得较高的分类正确率。

The automatic target recognition procedure of synthetic aperture radar（SAR） generally includes two steps,feature extraction and classifier training.Based on the development of deep convolutional neural networks,we present a new method of SAR target recognition.This method automatically learns the hierarchies of features from different targets,which means it avoids the nonnormalization caused by manual feature extraction.Then the transfer learning technology is applied to avert the occurrence of locally optimal solution and accelerate the training procedure.Finally we use the moving and stationary target acquisition and recognition database to verify our method.