Phase demodulation is the core of fringe projection profilometry systems. However, current U-Net-based phase demodulation approaches demonstrate deficiencies in global context propagation, adversely affecting wrapped phase extraction precision. To this end, this paper proposes a deep-learning-based model for single-shot wrapped phase extraction, named the full-scale connection and attention enhancement network (SEC-UNet3+). The network mitigates the limitations of the traditional U-Net architecture by introducing cross-layer full-scale connection and a feature integration module in the decoder, enabling efficient interaction between shallow detail features and deep semantic features. Unlike the skip connection strategy within the same-layer in U-Net, cross-layer full-scale connection can enhance the feature utilization. Additionally, a skip connection is embedded between the feature mapping layer and the output transformation layer in the squeeze and excitation module, preventing information loss during the feature calibration process. Compared to the U-Net model, the proposed method achieves an approximately 5% to 15% reduction in both the mean squared error and mean absolute error for phase extraction. The experimental results confirm that SEC-UNet3+ outperforms traditional Fourier transform and mainstream U-Net-based approaches in phase demodulation accuracy, proving particularly effective for single-shot wrapped phase retrieval in dynamic scenarios.