Time-frequency graph enhancement of communication signals based on generative diffusion model in low signal-to-noise ratio environment

Abstract In low signal-to-noise ratio (LSNR) environments, the processing of communication signals faces significant noise interference. Traditional signal enhancement methods often perform poorly under these conditions, making it difficult to effectively improve signal quality. To address this challenge, this paper proposes a dual-stage signal enhancement method based on an improved DiffBIR model (Diffusion-based Blind Image Restoration Model), aiming to efficiently enhance received signals through time-frequency diagrams, thereby improving signal quality in low SNR environments. The improved DiffBIR model combines the advantages of deep learning and diffusion processes, utilizing Inception and PFA modules to achieve adaptive signal recovery in the time-frequency domain. The Inception module provides a rich feature foundation for the PFA module through a multi-scale feature extraction mechanism, while the PFA module further enhances the accuracy of signal recovery by optimizing the weight distribution of signal regions. Experimental results show that under low SNR conditions, the proposed improved DiffBIR model significantly outperforms traditional signal enhancement methods, particularly in scenarios with very low SNR, where the enhancement effect is especially pronounced. This method offers an innovative solution for enhancing received signals, not only demonstrating high noise suppression capabilities but also better preserving the time-frequency characteristics of the signal. It has broad application prospects, particularly in fields such as communication, radar, and acoustic signal processing. The code and data supporting this research have been stored on GitHub, with the link being https://github.com/18291716943/demo.