A Resolution-Preserving Multi-Scale Network for Spectrogram-Based Semantic Segmentation of 5G-LTE Signals in Spectrum Sensing

In this paper, we propose a novel semantic segmentation model designed to enhance the accuracy of 5G New Radio (NR) and Long-Term Evolution (LTE) signal segmentation in spectrogram images for intelligent spectrum sensing. To address the challenge of preserving essential signal information in spectral feature learning, we developed Resolution-Preserving Multi-Scale Network (RPMSN), an effective deep network architecture for semantic segmentation. This model maintains input resolution and extracts features at multiple scales by exploiting two components: a resolution-preserving architecture and a multi-scale block, consequently making it particularly effective for segmenting wide band spectrograms. In addition to significantly outperforming several state-of-the-art architectures, including U-Net, U-Net++, DeepLabV3, DeepLabV3+, FPN, and PSPNet, RPMSN achieved impressive performance with a mean accuracy of 97.22%, mean IoU of 95.42%, mean F1 Score of 97.64% under the presence of different channel impairments while keeping the model size efficient at 14.5M parameters.