一种基于模糊滤波提高SAR自动目标识别平移不变性的方法

使用卷积神经网络去实现合成孔径雷达(synthetic aperture radar, SAR)自动目标识别(auto target recognition, ATR)成为了近年来的热点,但实际使用中存在的一个隐患问题是平移不变性的丢失,随着目标位置移动,系统输出也随之改变，从而造成错误识别。针对上述问题，提出了一种落实在模型层面的解决方法,通过对算法的改进,实现提升SAR ATR系统平移不变性,而无需数据增强。提出的模块易于移植到现有SAR ATR骨干网络中,且通过实测兼容良好,引入后不影响识别准确率,达到了与原网络近似相等甚至更高的精度。结果表明,所提出的算法不仅提升了系统的平移不变性,同时提升了系统的抗干扰能力。

Improving translation invariance of SAR automatic target recognition based on blur filtering method

Using convolutional neural network to realize synthetic aperture radar (SAR), auto target recognition (ATR) has become a hot spot in recent years.However, a hidden problem in practical use is the loss of translation invariance. As the target moves horizontally, vertically or both, the output of the system changes accordingly, which leads to misidentification. A practical solution is proposed at the model level. Through the improvement of the model algorithm, the translation invariance of the SAR ATR system is significantly improved, and there is no need for data augmentation. The proposed module is easy to be inserted into the existing SAR ATR backbone network, which is well compatible, has effect on recognition accuracy and achieves an approximate or slightly higher accuracy than the original network. Experiments demonstrate that the algorithm proposed not only improves the translation invariance of the system, but also improves the anti-interference ability of the system.