A speech enhancement method is proposed in this paper to address the issues of susceptible time-frequency features and insufficient local detail retention in complex noise environments. The proposed method is based on multi-scale receptive field and frequency-domain channel time-frequency feature fusion. To overcome the limitation of a single receptive field caused by fixed convolution kernels, a progressive multi-stage feature awareness module is introduced. This module aggregates contextual information from different scales, enhancing the perception of speech details and noise variations. Subsequently, a global frequency domain convolution module and parallel dual attention aggregation module are utilized to perform global convolution in the frequency domain. This compensates for the lack of global modeling and avoids dependence on noisy phases for reconstruction. At the same time, enhance the focus on key time-frequency areas and help to model long-distance dependencies. In the decoding phase, the phase direct prediction strategy is adopted, and the phase estimation is obtained through parallel regression of the real part and the imaginary part, combined with atan2 operation, to reduce the dependence on noisy phase information, thus helping to reduce phase artifacts. Finally, experiments were conducted on the VoiceBank+DEMAND dataset. The results show that the proposed network achieves a PESQ score of 3.52 and a STOI of 96%. It shows that the network can effectively improve the overall quality of voice signals, and has good noise reduction ability. We selected TIMIT speech that did not participate in the training and a variety of non noise to conduct cross dataset generalization experiments. The results show that the model has good generalization ability and robustness.In addition, a cross-domain test set constructed using clean Mandarin speech from AISHELL-1 and real-world environmental noise from RealMAN further validates the applicability of the proposed method to practical real-world acoustic scenarios.