DeGS-Net: Interference-Resilient Spectrum Sensing via Geometric-Semantic Consensus

Abstract

Accurate spectrum sensing is critical for optimizing the coexistence of radar and communication systems in dynamic wireless environments. While deep learning has emerged as a promising solution existing semantic segmentation models often face a critical trade off where high precision architectures are too resource intensive for real time applications while lightweight models sacrifice accuracy especially in noisy conditions. To address these challenges this paper introduces DeGS-Net a novel end to end framework designed for both high accuracy and computational efficiency. Its core innovation centers on a synergistic multi task paradigm where an auxiliary denoising task is strategically employed as a powerful regularization mechanism to force the shared encoder to learn noise invariant features. Through comprehensive experiments on a synthetic dataset containing 5G new radio, long term evolution and radar signals DeGS-Net demonstrates superior performance achieving a state of the art mean intersection over union of 64.29%. Furthermore it proves to be highly efficient delivering this high accuracy with only 1.51M parameters representing a significant reduction in complexity compared to other high performance models. With a compact parameter footprint DeGS-Net offers a highly effective and scalable solution for intelligent spectrum sensing in future wireless systems.