Deep Spectrum Sensing via Residual One-For-All Group Convolutions and Hierarchical Auxiliary Fusion

Abstract

The paradigm shift toward sustainable spectrum utilization in future wireless architectures requires spectrum sensing methods that balance fine-grained characterization of complex spectral scenes with minimal computational overhead. To this end, we propose OFA-UNet, a cost-efficient model for segmenting various signals (such as 5G, LTE, and radar) in time–frequency spectrograms. OFA-UNet introduces two core components to jointly optimize accuracy and architectural efficiency: (i) a One-For-All Group Convolution (OFA-GConv) module that suppresses structural filter redundancy to compress model size, and (ii) a Hierarchical Auxiliary Fusion (HAF) module leveraging a local-convolution and parameter-free SimAM-based Dual-Context Aggregation (DCA) block for improved gradient and multi-scale aggregation. OFA-UNet achieves 97.2% mAcc and 94.0% mIoU using just 0.8M parameters and a 1.52 ms inference time. Notably, it surpasses the baseline by 4.1% mIoU while reducing parameters by $18.0\times $ and GFLOPs by 50.9%, facilitating seamless edge deployment.