RIME-Net: A Physics-Guided Unpaired Learning Framework for Automotive Radar Interference Mitigation and Weak Target Enhancement

TL;DR

RIME-Net is a new deep learning framework that improves radar performance by reducing interference and enhancing weak targets without needing paired training data.

Contribution

The paper introduces RIME-Net, a physics-guided unpaired learning framework for joint interference mitigation and weak target enhancement in automotive radar.

Findings

RIME-Net outperforms existing methods in SINR, recall, and structural similarity.

The framework effectively suppresses interference while preserving signal integrity.

Experiments show robust performance across diverse datasets and environments.

Abstract

With the widespread deployment of automotive millimeter-wave radars, mutual interference and broadband noise severely degrade the signal-to-noise ratio (SNR) of range–Doppler (RD) maps, leading to the loss of weak targets. Existing deep learning methods rely on difficult-to-obtain paired training samples and often cause excessive target smoothing due to a lack of physical constraints. To address these challenges, this paper proposes RIME-Net, a physics-guided unpaired learning framework designed to jointly achieve radar interference mitigation and weak target enhancement. First, based on a cycle-consistent adversarial architecture, we designed the Interference Mitigation Network (IM-Net). IM-Net integrates spectral consistency loss and identity mapping constraints, learning a robust mapping from the interference domain to the clean domain without paired supervision, effectively suppressing low-rank interference and preserving signal integrity. Second, to recover target details attenuated during denoising, we propose the saliency-aware Target Enhancement Network (TE-Net). TE-Net combines multi-scale residual blocks and channel-spatial attention mechanisms, selectively enhancing weak target features based on saliency priors. Extensive experiments on diverse datasets show that RIME-Net significantly outperforms existing supervised and model-driven methods in terms of SINR, recall, and structural similarity, providing a robust solution for reliable radar perception in complex electromagnetic environments.

Figures