Multimodal recommender systems enhance collaborative filtering by integrating visual and textual features, yet confront two critical limitations: (1) The lack of an effective mechanism to adaptively adjust modality fusion weights, which limits their ability to accurately capture user preferences; and (2) neglected differences in representation granularity across modalities lead to suboptimal cross-modal semantic alignment. To address these challenges, we propose CRLMRec (Contrastive Representation Learning for Multimodal Recommendation) with dual innovations: (1) A dynamic multi-weight cross-modal fusion (DyMFuse) module adaptively adjusts fusion weights across modalities to effectively capture user preferences in each modality. (2) A Contrastive Hierarchical Cross-modal Alignment (CHCA) module aligns multimodal representations by maximizing mutual information between different modality views of the same item in both fine- and coarse-grained feature spaces. Comprehensive empirical evaluations on two public datasets validate the effectiveness of CRLMRec, achieving performance improvements of 11. 20% to 14. 34% compared to state-of-the-art baselines.