This paper introduces FRED-Lite, an advanced lightweight deep learning network with a convolutional architecture designed for wireless signal segmentation in time-frequency occupancy spectrograms. It aims to balance accuracy, processing speed, and model compactness, making it suitable for integration into embedded devices and resource-constrained systems. FRED-Lite employs a full-resolution encoder architecture, a boundary refinement module in the decoder, and an input grouped multi-kernel extractor module. This design efficiently captures spectral information, enhancing segmentation accuracy by extracting both local and global features for robust wideband spectrogram analysis under diverse signal and channel conditions. Experimental results demonstrate that FRED-Lite achieves an optimal trade-off between performance and computational complexity compared to state-of-the-art models, including U-Net, U-Net++, SegFormer, DeepLabV3+, SRNet, and RPMSN. With only 4.7M trainable parameters and an inference speed of 39.5 ms per image, the model achieves a mean accuracy of \(92.75\%\) , a mean intersection over union of \(86.59\%\) , a mean F1-score of \(92.81\%\) , and a mean precision of \(92.90\%\) , FRED-Lite presents a promising solution for intelligent spectrum sensing in next-generation wireless communication systems with limited storage and processing capabilities.