With the rapid development of social media platforms, the spread of multimodal fake news poses a serious threat to social stability. Existing detection methods suffer from two major limitations: single-granularity feature extraction that over-relies on abstract features from pre-trained models or is confined to shallow statistical features; and imperfect cross-modal fusion mechanisms that fail to effectively capture image-text semantic consistency. To address these challenges, we propose a multimodal fake news detection method based on Multi-Granularity Feature Collaborative Learning (MGFCL). The method innovatively constructs a multi-granularity feature extraction framework spanning from abstract semantics to contextual awareness features, achieves complementary fusion of different granularity features through concatenation, and employs attention mechanisms for cross-modal adaptive fusion. Experimental results on the Weibo dataset show that our method significantly outperforms existing baseline models, and validates the effectiveness of each component through ablation experiments. This study provides a novel feature fusion paradigm for multimodal fake news detection and establishes a solid foundation for future research in feature optimization and model improvement.