Fourier Neural Operator (FNO) has emerged as a powerful paradigm for solving partial differential equations (PDEs) by learning discretization-invariant solution operators that effectively capture global flow dependencies. However, due to the global nature of Fourier basis functions and the truncation of high-frequency modes, standard FNO models often struggle to resolve fine-grained local features and discontinuities, limiting their fidelity in capturing detailed flow structures. To address this limitation, this study introduces a hybrid CNN-FNO architecture for airfoil flow field prediction. The proposed framework employs a two-stage prediction structure. The first-stage FNO produces a coarse flow field prediction, while the parallel FNO and CNN branches in the second stage learn the global and local components of the prediction residual, respectively. A key innovation is the integration of a CBAM-based adaptive fusion mechanism that dynamically weights the contributions of both branches, enabling the model to synergize global coherence with local precision during the learning process. The model was evaluated on a supersonic airfoil flow dataset and a transonic NACA-Euler flow dataset. Numerical experiments demonstrate that this hybrid approach achieves superior predictive accuracy, reducing the overall relative L2 errors by 14.37% and 29.11% on the two datasets, respectively, compared with the standard FNO. Notably, this accuracy improvement incurs only a modest additional training cost, while the parameter count remains nearly unchanged. Furthermore, the proposed model offers a computational speedup of 3 to 5 orders of magnitude over traditional CFD solvers. By effectively integrating spectral information with local convolutional receptive fields, the proposed architecture significantly enhances the prediction of Mach number and pressure-related flow fields, offering a robust and efficient surrogate model for aerodynamic design optimization.