Time-Frequency Image Enhancement for Low SNR Communication Signals Using Generative Diffusion Models

In low signal-to-noise ratio (LSNR) environments, communication signals are severely interfered by noise, rendering traditional enhancement methods ineffective. This study proposes a dual-stage signal enhancement framework based on an improved DiffBIR model, integrating deep learning and diffusion processes. By incorporating the Inception module for multi-scale feature extraction and the Pixel Fusion Attention (PFA) module for key region highlighting, the model achieves enhanced signal recovery in the time-frequency domain. Experimental results demonstrate that the improved DiffBIR model outperforms traditional methods in both noise suppression and preservation of time-frequency characteristics, with broad application prospects in communication, radar, and acoustic processing.