基于深度学习的SAR目标识别技术研究

Research on SAR target recognition based on deep learning

对于SAR图像而言,二维图像的维度要远大于一维距离@像的维度。因此,无法直接对高维数据进行处理。为了解决该问题,我们提出了一种利用深度学习技术对SAR图像特征进行离线学习的方法。深度学习是一种通过多层结构进行机器学习的神经网络模型,直接从原始的信息中学习特征。深度学习提出后,在图像识别、语音识别以及目标侦测等许多领域都得到了广泛的应用,并取得了不错的效果。深度学习在不同应用中采取了不同的结构,其中,卷积神经网络(Convolutional Neural Network,CNN)因其可以直接处理二维数据而不破坏其拓扑性,所以常用于图像识别问题。文章针对SAR图像的目标识别问题,设计了一种表现良好的卷积神经网络模型;针对网络模型搭建中模型修改导致的重复训练问题,设计了一种基于NET2NET的迁移学习方法,有效降低了网络训练时间。

For Sar Images, the dimension of two-dimensional image is much larger than the dimension of one-dimensional range image. Therefore, high-dimensional data can not be processed directly. In order to solve this problem, we propose an off-line learning method for SAR image features using the depth learning technique. Deep learning is a neural network model of machine learning through multi-layer structure, learning features directly from original information. Deep learning has been widely used in many fields such as image recognition, speech recognition and target detection, and achieved good results. Deep learning takes on different structures in different applications, and Convolutional Neural Network(CNN) is often used in image recognition problems because it can process 2d data directly without destroying its topology. In this paper, a good performance convolutional neural network model is designed for SAR image target recognition, and a migration learning method based on NET2NET is designed for the repetitive training problem caused by model modification in network model building, the network training time is reduced effectively.