<p>Speech enhancement in complex acoustic environments is a critical technique for improving speech recognition accuracy and communication quality. Existing deep learning-based speech enhancement models suffer from insufficient global modeling capability of key speech components in the frequency domain, as well as significant redundancy in deep feature representations, making it difficult to simultaneously achieve high computational efficiency and strong discriminative performance. To address these issues, this paper proposes a speech enhancement model termed FSCC-MP-SENet, which integrates frequency-domain channel attention and feature compression to optimize MP-SENet, thereby improving both model performance and computational efficiency. Specifically, a Frequency-domain Channel Attention Network (FCANet) is introduced to enhance the modeling of key frequency-domain features, while a low-parameter convolution module with channel reconfiguration (SCConv) is incorporated to construct joint spatial-channel feature representations, structurally strengthening the extraction of core speech information. Experimental results on the VoiceBank-DEMAND dataset demonstrate that FSCC-MP-SENet achieves significant improvements of 0.15 and 0.02 in PESQ and STOI, respectively. Furthermore, compared with state-of-the-art models such as SEGAN, MetricGAN, ForkNet, and MTFAA, the proposed FSCC-MP-SENet consistently attains the best performance across four objective evaluation metrics, namely PESQ, STOI, SSNR, and COVL, exhibiting stable performance and strong robustness under low signal-to-noise ratio conditions, low-frequency noise, and diverse complex noise scenarios. These results indicate that FSCC-MP-SENet effectively enhances speech clarity and intelligibility, providing an efficient and robust solution for speech enhancement in complex environments.</p>