Lightweight Residual Networks with Ghost Blocks for Efficient Direction-of-Arrival Estimation

Accurate and computationally efficient direction-of-arrival (DoA) estimation remains a fundamental requirement in wireless communications, radar, and array signal processing. This paper presents novel residual network, a deep learning architecture that incorporates ghost blocks and squeeze-and-excitation (SE) modules into the ResNet framework to address the trade-off between accuracy and complexity. The introduction of ghost blocks reduces parameters and floating-point operations by approximately 50%, while SE modules improve feature discrimination with negligible overhead. Experimental evaluations demonstrate that the proposed architecture achieves high estimation accuracy, with F1 scores reaching 96.80% in single-signal scenarios and 89.21% with five simultaneous signals, alongside mean absolute errors between 0.016° and 0.161°. The proposed approach thus provides a balanced solution that combines precision with computational efficiency, supporting real-time deployment.