Polynomial Phase Signal Denoising Connecting Semantic Information Based on Deep Neural Networks

Abstract

This paper considers the problem of polynomial phase signal (PPS) denoising. To prove that proper use of semantic information can further improve the denoising performance based on deep neural networks, we propose an architecture combining the segmentation network and the denoising network. The vision semantic information is extracted from the segmentation network first. Then, that information connecting the time-frequency representation of noisy signal are fed into the denoising network for reconstructing signal. To effectively apply the semantic information, three connection strategies and the corresponding lower bound are presented and compared. The proposed method does not require the pre-identification of signal noise conditions and is suitable for a wide range of Signal-to-Noise-Ratio (SNR) scenarios. Simulation results demonstrate that the F1 scores of the spectrum segmentation results are over 0.98 and the proposed method connecting vision semantics for PPS denoising tasks outperforms the baseline and state-of-the-art architectures, when the SNR is larger than -8dB.