A multimodal recommendation model that integrates high-order semantic enhancement with adaptive modal feature fusion (MMSAF) is proposed. The model employs graph convolutional neural network (GCNN) to perform high-order semantic enhancement, enabling the capture of deeper associations between users and items, thereby more accurately reflecting users' complex preferences. The effectiveness and applicability of this model are validated. An adaptive modality fusion mechanism is introduced to dynamically adjust weights of modal features based on their relative importance in different contexts, enabling flexible adaptation to diverse user preferences. Experimental results demonstrate that MMSAF outperforms existing mainstream methods across multiple benchmark datasets in terms of recommendation accuracy and generalization capability.