错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Recognizer Embedding Diffusion Generation for Few-Shot SAR Recognization

  • Ying Xu,
  • Chuyang Lin,
  • Yijin Zhong,
  • Yue Huang,
  • Xinghao Ding

摘要

Synthetic Aperture Radar (SAR) has become a research hotspot due to its ability to identify targets in all weather conditions and at all times. To achieve satisfactory recognition performance in most existing automatic target recognition (ATR) algorithms, sufficient training samples is essentially owned. However, only few data for real-world SAR applications are generally available, arriving at efficient identification with few SAR training samples remains a formidable challenge that needs to be addressed. The success of denoising diffusion probabilistic model (DDPM) has provided a new perspective for few-shot SAR ATR. Consequently, this paper proposes the Recognizer Embedding Diffusion Generation (REDG), a novel approach for few-shot SAR image generation and recognition. REDG mainly consists of a generation module and a recognition module. The former is responsible for generating additional data from a limited SAR dataset by means of DDPM, to enhance the training of the recognizer. The latter is designed specifically to optimize data generation, which further enhances the recognition performance using the generated data. Extensive experiments conducted on three different datasets have demonstrated the effectiveness of proposed REDG, which represents a significant advancement in improving the reliability of SAR recognition systems, especially in scenarios lacking sufficient data.