Accurate Deeplofargram: Accurate dynamic dim frequency line recovery for underwater acoustic target via waveform...

Accurate estimation of frequency lines is crucial for correctly classifying and identifying underwater sources in remote sensing. However, the dynamic dim line recovery still faces challenges caused by low signal-to-noise ratio (SNR), uncertain frequency variations, and unpredictable birth and death time points. To address these issues, rather than solely focusing on frequency lines as conducted by existing methods, we propose a waveform diffusion model focusing simultaneously on the background noise as well. Concretely, in the forward process, noisy signals with arbitrary SNRs can be generated by progressively adding Gaussian noise. Then, in the reverse process, we train the model by predicting and removing the added noise. In this process, the non-Gaussian background noise can still be potentially modeled by the multivariate Gaussian noise through enough iterations. Moreover, to avoid accuracy degradation due to the loss of phase information, we specifically design a transfer probability kernel for waveform signals instead of lofargram. Extensive experiments demonstrate that the proposed model achieves the best results on both simulated and real-world datasets, handling both fluctuating and straight frequency lines. Additionally, our method can also identify the birth and death time points with low SNRs.