Abstract: To address the issues of high cost and high computational complexity in traditional array radar systems for Direction of Arrival (DOA) estimation, this paper proposes a DOA estimation method that combines the amplitude-phase detection of conventional array antennas with the amplitude detection of metalens antennas. The front-end of the proposed DOA estimation system consists of multiple conventional amplitude-phase detection units and metalens-based amplitude detection units. Based on this front-end architecture, a dual-branch neural network named Fusion-SE-ResNet is designed. This network can separately process complex-valued and real-valued signals, incorporating residual connections and channel attention mechanisms to achieve deep feature extraction and adaptive fusion of heterogeneous signals.In dual-target tests with a Signal-to-Noise Ratio (SNR) of 10 dB, the Fusion-SE-ResNet model achieves an angle estimation accuracy of over 99%. Under a low SNR of 5 dB, it maintains an accuracy above 92%, whereas traditional DNN models only reach 80%, demonstrating superior robustness. In three-target scenarios, Fusion-SE-ResNet effectively resolves targets with angular separations of 5°and 27°.Thanks to the low cost of metalens and amplitude detection units, this solution achieves higher-precision DOA estimation with fewer conventional amplitude-phase detection units, reducing overall system cost while improving performance.