基于散射信息和元学习的SAR图像飞机目标识别

Scattering Information and Meta-learning Based SAR Images Interpretation for Aircraft Target Recognition

SAR图像由于数据获取难度大,样本标注难,目标覆盖率不足,导致包含地理空间目标的影像数量稀少。为了解决这些问题,该文开展了基于散射信息和元学习的SAR图像飞机目标识别方法研究。针对SAR图像中不同型号飞机空间结构离散分布差异较大的情况,设计散射关联分类器,对飞机目标的离散程度量化建模,通过不同目标离散分布的差异来动态调整样本对的权重,指导网络学习更具有区分性的类间特征表示。考虑到SAR目标成像易受背景噪声的影响,设计了自适应特征细化模块,促使网络更加关注飞机的关键部件区域,减少背景噪声干扰。该文方法有效地将目标散射分布特性与网络的自动学习过程相结合。实验结果表明,在5-way 1-shot的极少样本新类别识别任务上,该方法识别精度为59.90%,相比于基础方法提升了3.85%。减少一半训练数据量后,该方法在新类别的极少样本识别任务上仍然表现优异。

The sample scarcity issue is still challenged for SAR images interpretation.The number of geospatial targets related images is constrained of the SAR images interpretation ability of data acquisition,sample labeling,and the lack of target coverage.Our SAR-ATR method is demonstrated based on scattering information and meta-learning.First,the discrete distribution of the spatial structure of different types of aircraft is quite different in SAR images.An associated scattering classifier is designed to guide the network to learn more discriminative intra-class and inter-class feature descriptions.Our proposed classifier facilitates the modeling of discrete degree of the aircraft target quantitatively and balance the weights of sample pairs dynamically through the differentiated analysis of different target discrete distributions.In addition,an adaptive feature refinement module is designed to optimize the network cohesion for the key parts of the aircraft and reduce the interference of background noise.The proposed method integrates the target scattering distribution properties to the network learning process.On 5-way 1-shot emerging categorized recognition task involved only few samples,our experimental results demonstrate that the recognition accuracy of this method is 59.90%,which is 3.85%higher than the benchmark.After reducing the amount of training data by half,the proposed method is still competitive on the new category of few-shot recognition tasks.