LESS-Net: Lightweight and Efficient Wideband Spectrum Segmentation Network

Conventional deep learning-based spectrum sensing approaches rely on over-parameterized models for high accuracy, which poses many challenges when deploying them in real-time scenarios with limited resources. In this work, we present LESS-Net, a lightweight and efficient model built upon improvements to the standard UNet architecture. The proposed approach is based on semantic spectrum segmentation, which allows multiple signals to be detected, localized, and classified simultaneously. Experimental results show that LESS-Net achieves excellent performance in localizing and recognizing different signals in the time-frequency space with fewer parameters and computations, making it suitable for practical applications.