Enhanced Residual Networks with Squeeze-and-Excitation Blocks for Multi-Signal DoA Estimation

Accurate direction of arrival (DoA) estimation is a fundamental challenge in wireless communications, radar, and array signal processing. In this paper, we propose a novel deep learning architecture based on squeeze-and-excitation residual networks (SE-ResNet) to enhance DoA estimation performance in challenging multi-signal and low signal-to-noise ratio (SNR) scenarios. By introducing adaptive channel attention mechanisms within the residual blocks, the proposed SE-ResNet dynamically recalibrates feature maps to emphasize the most informative spatial features. This enables significant improvements in multisource DoA estimation accuracy while maintaining computational efficiency suitable for real-time applications. Extensive experiments demonstrate that SE-ResNet consistently outperforms a standard ResNet baseline across various numbers of simultaneous signals and SNR conditions. Notably, the proposed approach achieves substantial reductions in mean angular error and boosts classification F1 scores with only a marginal increase in computational complexity.