任务感知的多尺度小样本SAR图像分类方法  

Task-aware Few-shot SAR Image Classification Method Based on Multi-scale Attention Mechanism

作　　者：张睿 王梓祺 李阳 王家宝 陈瑶 ZHANG Rui;WANG Ziqi;LI Yang;WANG Jiabao;CHEN Yao(Command and Control Engineering College,Army Engineering University of PLA,Nanjing 210007,China)

机构地区：[1]陆军工程大学指挥控制工程学院,南京210007

出　　处：《计算机科学》2024年第8期160-167,共8页Computer Science

基　　金：江苏省自然科学基金(BK20200581)。

摘　　要：针对SAR图像分类时存在的带标注样本较少的问题,提出了一种任务感知的多尺度小样本SAR图像分类方法。为了能够充分挖掘局部特征并关注具体任务下的关键局部语义patches,引入了两种有效的注意力机制,获得了更加高效且丰富的特征表示。首先,在特征提取阶段使用互补注意力模块(CSE Block),关注原始特征中不同语义部分的显著特征,从被抑制的特征中提取次级显著特征并与主要显著特征融合,得到更加高效且丰富的特征表示。随后,利用自适应情景注意力模块(AEA Block)获得整个任务中的关键语义patches,增强任务间的区分信息,提升小样本SAR图像分类任务的精度。结果表明,在SAR图像分类标准数据集MSTAR上,5-way 1-shot任务分类精度相较于次优方法精度提升了2.9%,并且该方法在两项任务中的运行时间与其他度量学习方法相比水平相当,未额外增加过多的计算资源,验证了其有效性。Aiming at the problem of the lack of labeled samples in SAR image classification,this paper proposes a task-aware few-shot SAR image classification method based on multi-scale attention mechanism.In order to fully mine local features and focus on the key local semantic patches under specific tasks,this paper introduces two effective attention mechanisms to obtain more efficient and rich feature representation.First,in the feature extraction stage,the complemented squeeze-and-excitation attention block(CSE Block)is used to focus on the salient features of different semantic parts of the original features.It can extract secon-dary salient features from the suppressed features and merge them with the main salient features,which can obtain more efficient and rich feature representation.Subsequently,an adaptive episodic attention block(AEA Block)is used to obtain key semantic patches in the entire task,which can enhance the differentiated information between tasks and improve the accuracy of SAR image classification tasks.The results show that the classification accuracy of the 5-way 1-shot task is 2.9%higher than that of the sub-optimal task on the SAR image classification standard MSTAR dataset.In the two tasks,the runtime of the proposed method is the same as other metric-learning methods,without additional excessive computing resources,which verifies its effectiveness.

关 键 词：多尺度注意力机制 小样本学习 SAR图像分类 度量学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]