| 摘要: |
| 合成孔径雷达(SAR)目标自动识别(ATR)技术的研究已经引起了广泛关注。然而，在实际应用中，雷达在单次观测中难以捕获所有方位角下目标的数据，且SAR目标图像特征对方位角高度敏感，导致现有的SAR-ATR模型经常出现在未观测方位角域数据上性能显著下降现象。当新的观测数据以流的形式出现时，仅依赖新观测数据对现有模型进行训练，将导致模型对已学习的方位角下SAR目标特征产生灾难性遗忘。因此，本文通过引入Fisher信息矩阵调节的正则项来保护对识别任务贡献显著的模型参数，并利用核心集减小推理误差，构建一种方位角自适应SAR目标识别连续学习模型。在MSTAR数据集上进行实验，实验结果表明，方位角自适应SAR-ATR模型能够在线学习不同方位角下SAR目标数据，不断适应不同方位角下SAR目标特征变化，识别精度得到有效提高。 |
| 关键词: 合成孔径雷达 自动目标识别 方位角域 连续学习 Fisher信息矩阵 |
| DOI： |
| 分类号:TN958 |
| 基金项目:国家自然科学基金(No.62071440) |
|
| Azimuth-Adaptive SAR Target Recognition Based on the Fisher Information Matrix |
|
|
|
|
| Abstract: |
| Synthetic Aperture Radar (SAR) Target Automatic Recognition (ATR) has garnered significant attention due to its potential applications. However, in real-world, it is difficult for radar to capture SAR target data across all azimuth angles in one observation. Additionally, SAR target image features are highly sensitive to the azimuth angle, leading to the performance degradation of existing SAR target recognition models when facing unknown azimuth domain data. When new SAR target data arrive in the form of a stream, only relying on new data to retrain the existing model will lead to catastrophic forgetting of the learned SAR target features in azimuth. To address these challenges, we propose an azimuth-adaptive continuous learning model for SAR target recognition, which incorporates a regularization term based on the Fisher information matrix to protect critical model parameters, and utilizes a core set to reduce inference error. Experimental results on the MSTAR dataset show that the proposed model can effectively learn SAR target data under different azimuth angles online, continuously adapt to changes in target characteristics, and significantly improve recognition accuracy. |
| Key words: synthetic aperture radar automatic target recognition azimuth domain continuous learning Fisher information matrix |