Detecting radar targets swarms in range profiles with a partially complex-valued neural network

Correctly detecting radar targets is usually challenged by clutter and waveform distortion. An additional difficulty stems from the relative proximity of several targets, the latter being perceived as a single target in the worst case, or influencing each other’s detection thresholds. The negative impact of targets proximity notably depends on the range resolution defined by the radar parameters and the adaptive threshold adopted. This paper addresses the matter of targets detection in radar range profiles containing multiple targets with varying proximity and distorted echoes. Inspired by recent contributions in the radar and signal processing literature, this work proposes partially complex-valued neural networks as an adaptive range profile processing. Simulated datasets are generated and experiments are conducted to compare a common pulse compression approach with a simple neural network partially defined by complex-valued parameters. Whereas the pulse compression processes one pulse length at a time, the neural network put forward is a generative architecture going through the entire received signal in one go to generate a complete detection profile.

💡 Research Summary

The paper tackles the challenging problem of detecting multiple, closely spaced radar targets—often referred to as “target swarms”—in range profiles (RPs) that suffer from clutter, sidelobe interference, and waveform distortions. Traditional processing pipelines rely on a matched filter (MF) followed by a cell‑averaging constant false‑alarm rate (CA‑CFAR) detector. While optimal for a single target in white Gaussian noise, the MF‑CA‑CFAR chain degrades when a weak target is masked by the sidelobes of a stronger one, when targets are densely packed, or when the received waveform is altered (e.g., reduced bandwidth, plasma‑coated stealth targets).

To overcome these limitations, the authors propose a partially complex‑valued neural network that processes the entire received IQ signal in a single forward pass, producing a detection profile of the same length as the RP. The architecture is hybrid: the first linear layer and its activation are complex‑valued, using the modReLU function (ReLU applied to the magnitude plus a bias, preserving the original phase). The complex output’s real part is then fed to a conventional real‑valued linear layer, ending with a sigmoid that maps each range bin to a probability in