AC-UNet: An Enhanced UNet for Multi-Class Signal Segmentation

Signal detection and identification play a critical role in ensuring communication security and improving spectrum utilization efficiency. With the rapid advancement of wireless communication technologies, electromagnetic environments have become increasingly complex, characterized by diverse signal types and high signal density, which poses significant challenges for signal detection and identification. However, the existing methods still have deficiencies and cannot simultaneously achieve high-precision multi-class signal detection and identification. To address this limitation, this paper proposes a novel timefrequency image (TFI) segmentation network named AC-UNet for multi-class signal segmentation. The model based on the UNet architecture with skip connections, integrates an Atrous Spatial Pyramid Pooling module and a Convolutional Block Attention Module into the encoder. Through this design, the model effectively extracts and integrates multi-scale contextual features, thereby enhancing its capacity for deep feature representation and ultimately leading to significant improvements in multi-class signal segmentation accuracy. Experimental results demonstrate that the proposed model significantly outperforms the baseline methods in segmentation accuracy, achieving improvements of 9.5% in mean Dice and 13.2% in mean Intersection-over-Union.