ABSTRACT

Target discrimination is an important part of the synthetic aperture radar automatic target recognition (SAR ATR). Nowadays, convolutional neural network (CNN) has been used in SAR ATR successfully. However, training CNN requires large amounts of labelled data and collection of the labelled SAR data is expensive and time demanding. It may yield overfitting when directly applying CNN to the SAR target discrimination with the limited labelled SAR data. To tackle this problem, we design a semi-supervised SAR target discrimination framework consisting of the classification network and the reconstruction network. In addition to the labelled SAR data, the unlabelled SAR data are also used to help the whole network better extract generalized feature for classification. Moreover, to make the learned feature more discriminative, the feature constraint based on Kullback-Leibler (KL) divergence is introduced to minimize the distribution divergence between the training and test data feature representations. Experimental results on the miniSAR data show the effectiveness of the proposed method.

摘要

目标识别是合成孔径雷达自动目标识别（SAR ATR）的重要组成部分。如今，卷积神经网络（CNN）已成功用于SAR ATR。但是，训练CNN需要大量的标记数据，而标记SAR数据的收集既昂贵又费时。当使用有限的标记SAR数据将CNN直接应用于SAR目标判别时，可能会产生过度拟合的情况。为了解决这个问题，我们设计了一个由分类网络和重建网络组成的半监督SAR目标判别框架。除了标记的SAR数据外，未标记的SAR数据还用于帮助整个网络更好地提取广义特征进行分类。此外，为了使学习的功能更具区分性，引入了基于Kullback-Leibler（KL）散度的特征约束，以最小化训练和测试数据特征表示之间的分布散度。miniSAR数据的实验结果证明了该方法的有效性。