Abstract: In order to improve the SAR (synthetic aperture radar) target recognition performance, especially for variant target recognition performance, this paper proposes a sparse coding tree framework for radar target recognition based on the sparse representation recognition method. Sparse coding tree is a classification tree that uses node-specific dictionaries and classifiers at each node. In the training stage, this paper designs two types of classifiers respectively for target identification requirements and variant target recognition requirement. The root node is built to identify target type, while child nodes of the tree are built to identify variant target series. In the testing stages, the sparse coding tree recognizes the target type on the root node. Then according to the recognition result of the root node, the tree decides whether branching the test sample to a child node or outputting final recognition result. Target with variant type will branch to the child node trained for this specific target type group. In addition, target without a variant will simply output a final classification label. The experimental results on MSTAR (moving and stationary target acquisition and recognition) SAR image data sets show that the sparse coding tree improves the ability of variant target  recognition, while obtains similar target category recognition accuracy compared to the mainstream method.

Key words: SAR target recognition, variant target recognition；sparse coding tree, dictionary learning；sparse representation

摘要： 为了提高利用合成孔径雷达（synthetic aperture radar，SAR）图像对目标型号识别的能力，在稀疏表示识别方法的基础上，提出了一种树形框架稀疏编码的雷达目标识别方法。稀疏编码树是由多个节点构成的分类器，其上每个节点由不同识别需求的子分类器构成。在训练阶段，分别针对目标型号识别需求以及型号识别需求学习相应分类器，组成分类器的根节点和子节点。识别阶段在根节点位置完成对目标类别的判断，再根据根节点的判断结果，对存在型号变体的目标，在子节点上再对型号进行识别，最终输出目标的识别结果，而不存在型号变体的目标则直接输出识别结果。基于美国运动和静止目标获取与识别（moving and stationary target acquisition and recognition，MSTAR）计划录取的SAR图像数据集上的实验结果表明，树形结构在取得与主流方法相当的目标类别识别精度的前提下，提高了对目标型号的识别能力，同时能够准确输出目标类别识别结果。

关键词: SAR目标识别, 型号识别, 稀疏编码树, 字典学习, 稀疏表示