Advances in generative models have significantly improved the realism of synthetic content, raising increasing public concern over their potential misuse. Existing frequency-based approaches utilize limited spectral characteristics and often fail to capture discriminative patterns, resulting in poor generalization, especially when faced with images generated by unseen model architectures. To address the challenges of detecting AI-generated images across diverse generative models, we propose FANB-Net, a novel framework that enhances spectral discrimination and generalization by integrating global and local frequency cues. FANB-Net integrates two complementary branches: The global frequency perception branch leverages the Discrete Fourier Transform to capture holistic high-frequency artifacts across the entire image. The extracted representations are further refined using a Frequency-Awared Spatial-Channel Attention (FASCA) mechanism, which enhances the response to high-frequency information from both spatial and channel dimensions, and by an adaptive channel dropout strategy that improves sensitivity to global generation traces. Simultaneously, the local frequency analysis branch focuses on fine-grained frequency patterns across multiple scales and is enhanced by a Noise-injected High-frequency Boosting (NHB) module. Experiments on 32 synthetic image datasets show that FANB-Net achieves a 4.3% improvement in cross-domain generalization on unseen generation models and attains an AP of 97.4% on diffusion-generated images, outperforming state-of-the-art methods.