In the multimodal knowledge graph completion (MMKGC) task, improving performance hinges on effectively integrating structural information with multimodal data. Existing methods primarily model graph structures in isolation or simply aggregate single-modal features, failing to capture entity-specific structural semantics and cross-modal relational dependencies. We propose StructCog, a framework that collaboratively combines fine-grained multimodal fusion with structure-aware contrastive learning. Our Structural-Attentive Fine-grained Entity Fusion (SA-FEF) mechanism dynamically allocates relation-specific attention weights based on multimodal entity signatures, enabling context-aware embeddings that preserve local structural context while maintaining global semantic consistency. To further enhance structural discriminability, we introduce hierarchical contrastive learning, adaptively sampling hard negative samples to selectively amplify meaningful entity-relation patterns and effectively suppress spurious correlations. Extensive experiments on standard benchmarks demonstrate that our method consistently outperforms state-of-the-art baselines in maintaining structural consistency and resolving multimodal semantic conflicts. Additionally, the proposed approach shows strong generalization capabilities across different knowledge graph scales and modality combinations.