【摘要】 针对深度学习模型易出现灾难性遗忘的关键难点，提出了一种基于原型对比的合成孔径雷达(synthetic aperture radar, SAR)图像增量小样本目标检测算法-InFSAR(prototype contrast based incremental few-shot SAR object detection)。首先，采用基础数据集对检测器进行预训练，以构建初步的特征提取能力；其次，设计一种类原型表征生成模块，以构建一组能够代表数据内在特征的类原型。在增量学习阶段，设计一种混合类原型对比编码模块，以有效学习新类别与基础类别之间的区分性特征。此外，为缓解灾难性遗忘问题，引入类原型校准策略，使模型在类原型上的预测分布逐步逼近真实分布，从而保持对基础类别识别的稳定性。在小样本目标检测数据集SRSDD-v1.0上的实验结果表明，在5-shot设置下，InFSAR对船舶细粒度目标的检测精度达到46.5%。同时，该方法能够在无需访问基础类训练数据的情况下，实现对少量标注新类别的增量检测与识别。更多还原

【Abstract】 To address the critical challenge of catastrophic forgetting in deep learning models, this paper proposes a prototype contrast based incremental few-shot object detection algorithm for synthetic aperture radar(SAR) image,namely as InFSAR. Firstly, the detector is pre-trained on a base dataset to establish preliminary feature extraction capabilities. Secondly, a class prototype representation generation module is designed to construct a set of class prototypes that represent the intrinsic characteristics of the data. During the incremental learning phase, a mixed class prototype contrastive encoding module is designed to effectively learn discriminative features between new and base classes. Furthermore, to mitigate catastrophic forgetting, a prototype calibration strategy is introduced, guiding the model predictive distribution on class prototypes to gradually approximate the true distribution, thereby maintaining stability in recognizing base classes. Experimental results on the few-shot object detection dataset SRSDD-v1.0 show that under the 5-shot setting, InFSAR achieves a detection accuracy of 46.5% for fine-grained ship targets. Additionally,the proposed method can incrementally detect and identify new categories with limited annotations without requiring access to the base class training data.更多还原