Noise robust HRRP sequence recognition based on a deep unfolded go decomposition network

The performance of most high-resolution range profile (HRRP) recognition methods degrades dramatically under low-SNR conditions. Based on the scattering center model, an HRRP sequence power matrix consisted of multiple sequential power HRRPs can be decomposed into three components: a low-rank self term, a sparse cross term and a noise term. Notably, the low-rank self term contains most target structure signatures that are essential for recognition. This paper aims to achieve noise robust recognition by separating the self term from the HRRP sequence power matrix based on the low-rank decomposition theroy. Since the performance of the traditional low-rank decomposition algorithms heavily depends on manually selected parameters and needs numerous iterations, a deep unfolded Go Decomposition Network (GoDecNet) is proposed. Specifically, we improve and unfold the Go Decomposition (GoDec) algorithm into a network to estimate the self term and suppress the noise over the range profile. Additionally, to obtain more robust temporal-spatial features, we design a CNN-based module to extract fast-time structure features and introduce the diagonally-structured state space model to explore slow-time temporal correlations. Finally, a hybrid loss function is designed to train the network end-to-end and facilitate interaction between the modules. Experiments conducted on measured data and simulatied HRRP data demonstrate the superior performance of the proposed method under low-SNR conditions.