SSPWave: Integrated signal subspace projection wavelet-inspired network for HRRP denoising and recognition

Identifying non-cooperative targets based on HRRP is a critical and challenging task. To enhance HRRP recognition performance in harsh environments characterized by low signal-to-noise ratios (SNR), we innovatively proposed a noise-robust model that combines domain knowledge and time-frequency multi-resolution analysis, namely integrated signal subspace projection wavelet-inspired network (SSPWave). It cascades a fine-grained deep denoising model and a general recognition model. First, we attempt to integrate discrete wavelet transform (DWT) into the deep denoising model, systematically removing the high-frequency components corresponding to the noise layer by layer, while retaining the low-frequency components containing the main structure of the target on down-sampling process. Second, to reconstruct the high-frequency details required during up-sampling, we propose a signal subspace projection (SSP) module. Notably, SSP introduces the estimated SNR as prior, and facilitates waveform preservation through adaptive subspace projection. SSPWave achieves a balance between noise suppression and detail preservation with SNR-guided, demonstrating the flexibility and effectiveness in addressing various noise levels of HRRPs. We evaluated the model on two measured HRRP datasets, which exhibited advanced recognition robustness on several evaluation metrics. Most importantly, domain knowledge assistance and time-frequency multi-resolution analysis are validated as effective strategies for HRRP denoising and recognition tasks.