Multimodal recommendation improves performance by integrating visual and textual information, but it still faces two major challenges: modality noise interference and insufficient user preference modeling. To address these issues, a denoising and preference-aware multimodal recommendation model, DPRec, was proposed. A dual denoising mechanism was constructed to enhance feature quality. In the frequency domain, signal-level noise was adaptively filtered, while an ID-guided gating unit was introduced to suppress semantic-level redundancy, thereby improving the stability and expressiveness of multimodal representations. For structural modeling, both item-item and user-item graphs were incorporated to capture higher-order relationships and strengthen multimodal semantic propagation. For preference modeling, a dynamic user preference inference mechanism was designed. Long-term interest representations were combined with contextual information, and adaptive weights were assigned to different modalities to achieve personalized representation updates. In addition, a contrastive learning strategy was applied to align collaborative behavior representations and preference representations within a unified semantic space, which enhanced feature consistency and robustness. Experimental results showed that DPRec outperformed baseline methods on four public datasets, achieving an average improvement of approximately 4.22% in NDCG@20.