The effective exploitation of consensus and complementary information across multi-view data remains crucial for advancing multi-view representation learning. While existing shared and specific feature learning frameworks have demonstrated potential by disentangling shared and view-specific patterns, current approaches predominantly overlook the hierarchical nature of shared information among views, leading to suboptimal redundancy reduction in the fusion information. To address these limitations, we propose a novel multi-view representation learning with refined fusion information exploration approach that systematically explores multi-level shared features and view-specific characteristics through three distinct information channels: 1) high-level shared information across all views, 2) low-level shared information between view pairs, and 3) unique view-specific information. Specifically, we implement three specialized subnetworks per view to respectively capture these information categorized through innovative constraints: cross-view correlation analysis for pairwise low-level shared features, adversarial alignment for high-level common representations, and orthogonality regularization to ensure information exclusivity while minimizing redundancy. The proposed framework achieves superior feature disentanglement through a tripartite separation mechanism that maintains discriminability between shared and specific features, and hierarchical differentiation of shared information. Extensive experiment results on several public datasets justify the effectiveness of the proposed approach.