Abstract
Extracting micro-Doppler (m-D) curves from timefrequency distributions (TFD) often encounters difficulties due to signal overlap, component discontinuity, and high noise levels. These challenges limit the effectiveness of such extractions in accurately representing the underlying signal characteristics. Addressing these limitations, this paper introduces a deep learning approach — the Recurrent Self-Attention Separation Network (RSS Net). The network separates independent m-D curves by dividing the TFD of multi-component signals into distinct masks for each signal component. These masks can be integrated with traditional methods or directly used for such estimations. We have compared and analyzed the performance of both approaches under various conditions. The network's effectiveness has been validated through extensive testing on the m-D curves of spatial conical targets in complex scenarios. Our experimental results emphatically demonstrate the network's superior performance over existing methodologies.
Funding Information
  • the national key Research and Development Program of China (2021YFC3090402)
  • National Natural Science Foundation of China (62471497)
  • Postgraduate Scientific Research Innovation Project of Hunan Province (CX20230111)

This publication has 28 references indexed in Scilit: