The Swin transformer has recently attracted attention in medical image analysis due to its computational efficiency and long-range modeling capability, which enables the establishment of more distant relationships between corresponding voxels. However, transformer-based models split images into tokens, which results in transformers that can only model and output coarse-grained spatial information representations. To address this issue, we propose Recovery Feature Resolution Network (RFRNet), which enables the transformer to contribute with fine-grained spatial information and rich semantic correspondences. Furthermore, shifted window partitioning operations are inflexible, indicating that they cannot perceive the semantic information over uncertain distances and automatically bridge the global connections between windows. Therefore, we present a Weighted Window Attention (WWA) to automatically build global interactions between windows after the regular and cyclic shifted window partitioning operations for Swin transformer blocks. The proposed unsupervised deformable image registration model, named RFR-WWANet, senses the long-range correlations, thereby facilitating meaningful semantic relevance of anatomical structures. Qualitative and quantitative results show that RFR-WWANet achieves significant performance improvements over baseline methods. Ablation experiments demonstrate the effectiveness of the RFRNet and WWA designs.