SAR-LAM:面向小样本SAR目标识别的轻量化适应策略  

SAR-LAM:A Lightweight Adaptation Method Being Geared to Few-Shot SAR Target Recognition

作　　者：史松昊 王晓丹[1] SHI Songhao;WANG Xiaodan(Air Defense and Antimissile School,Air Force Engineering University,Xi’an 710051,China)

机构地区：[1]空军工程大学防空反导学院,西安710051

出　　处：《空军工程大学学报》2024年第3期103-111,共9页Journal of Air Force Engineering University

基　　金：国家自然科学基金(61876189,61703426,61806219)。

摘　　要：针对小样本学习中跨域迁移导致模型性能下降的问题,提出一种面向小样本SAR目标识别的轻量化适应策略(SAR-LAM)。该方法通过知识蒸馏预训练一个具有泛化性能的通用编码器,向其中嵌入一个只在少量目标域样本上进行训练的适应模块,而后将提取的特征映射到一个分辨性更高的空间内,最终以原型网络为基线对查询集样本进行分类。该适应策略以增加少量学习参数为代价,克服了数据分布差异导致模型迁移受限的困难,增强了模型在目标域提取特征的能力,在小样本条件下将SAR目标识别的准确率提升了至少1.93个百分点,较其他方法展现出一定的优越性。In view of the issue of model performance degradation caused by cross-domain transfer in few-shot learning,a lightweight adaptation strategy for few-shot SAR target recognition named SAR-LAM is proposed.This method is to utilize knowledge distillation for pre-training a generalized encoder and em-bedding an adaptation module trained only with very few target domain samples.The extracted features are then mapped into a more discriminative space,and finally,the query set samples are classified by tak-ing a prototypical network as the baseline.This adaptation strategy is to increase at a few cost in learning parameters,and by so doing,the limitations of model transfer caused by data distribution differences is overcome,improving the model’s ability to extract features in the target domain,and simultaneously im-proving the accuracy of SAR target recognition by at least 1.93 percentage points under few-shot condi-tions.And this adaptation strategy is superior in performance to the other methods.

关 键 词：SAR目标识别 跨域小样本学习 轻量化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]