D2SNet: A Denoising-to-Segmentation Network for Enhanced 5G-LTE Spectrum Awareness

This paper introduces D2SNet, a novel dual-task architecture that enhances spectrum sensing by synergistically integrating spectrogram denoising and segmentation. To address the critical challenge of identifying signals in noisy spectrograms, D2SNet’s denoising module introduces two architectural innovations: a dual-branch rotational learning strategy to learn orientation-invariant features and an edge-center swapping block to preserve sharp signal boundaries. By producing a refined spectrogram with enhanced visual clarity, D2SNet establishes a more reliable foundation for accurate segmentation of 5G and LTE signals. The model’s effectiveness derives from its decoupled two-stage training strategy, which optimizes the denoising and segmentation tasks independently to maximize accuracy. Simulation results demonstrate that D2SNet outperforms state-of-the-art models, achieving a mean IoU of 80.33% and a mean F1-score of 89.04% at 0 dB SNR. With only 13.6M parameters and a rapid inference time of 14.2 ms, D2SNet delivers a powerful and highly efficient solution for intelligent spectrum sensing in next-generation wireless systems.