Resolution-Preserving Multi-Scale Network for 5G-LTE Spectrogram-Based Spectrum Sensing

Abstract

This letter introduces PRMNet, an advanced deep model with a convolutional architecture for segmenting wireless signals in time-frequency occupancy spectrograms, designed to enhance spectrum sensing accuracy. PRMNet features a resolution-preserving architecture combined with a multi-scale feature extraction module to capture features across multiple resolution levels while preserving the original resolution of input spectrograms. This design facilitates an effective extraction of both local and global spectral features, thus robustly analyzing wideband spectrograms under diverse signal and challenging channel conditions. Experimental results reveal that PRMNet outperforms state-of-the-art models, such as FCN, U-Net, U-Net++, DeepLabV3+, and SegFormer, with a mean accuracy of 92.44%, a mean IoU of 87.6%, a mean F1 score of 93.35%, and a mean precision of 94.65%. Despite its compact architecture with only 14.5M parameters, PRMNet delivers remarkable performance, thus offering a promising solution for intelligent spectrum sensing in next-generation wireless communication systems.