A weak target detection method via joint target-background decomposition based on an improved Res-UNet

Weak-target detection in radar data acquired in complex environments is challenged by weak target returns, structured backgrounds, and false targets frequently observed in empty scenes. To improve discrimination between target returns and background-induced spurious responses, a Joint Target-Background Decomposition Detection Network (JTBD-Net) based on an improved residual U-Net (Res-UNet) is proposed. The network jointly uses in-phase/quadrature (I/Q) data and amplitude images as inputs and comprises target, background, and global discrimination branches. The target and background branches learn target-related and generalized-background features, respectively, so that background information can be explicitly retained and involved in the detection decision. By introducing background features as a discrimination reference, decomposition gating and differential fusion are employed to adaptively regulate target and background responses, enhancing weak-target features while suppressing background interference and false responses in empty scenes. Experimental results show that, at a signal-to-noise ratio of −18 dB, JTBD-Net improves the probability of detection by 11.4% over RadarFormer. In an empty-scene case with moderate-intensity false-target interference at −2.5 dB, it lowers the empty-scene false-alarm probability by 20.3% relative to UNet++. These results show that JTBD-Net improves weak-target detectability while effectively suppressing false alarms caused by complex backgrounds and false targets in empty scenes.