To address the challenges of low registration accuracy under complex deformations and limited generalization ability in existing medical image registration algorithms, a dual-stream registration network that integrates multi-scale residual learning with a multi-dilated perception Transformer (Dilformer) was proposed. First, a multi-scale residual learning block (MSR) was introduced to enhance feature representation during the dual-stream pyramid feature extraction stage. Then, the Dilformer module was designed to construct a heterogeneous receptive field interaction mechanism using multi-rate dilated convolutions, thereby improving the model’s global modeling capacity at low-resolution scales. Subsequently, a separable residual fusion block (SRF) was developed to effectively fuse multi-scale features and enhance the accuracy of the predicted deformation field. Finally, a multi-resolution loss function was introduced to supervise network training across multiple scales, further improving registration performance. Experimental results on the 3D brain MRI datasets LPBA40 and preprocessed IXI demonstrate that the proposed network achieves superior accuracy compared to state-of-the-art models. Specifically, on the IXI dataset, the proposed network achieves a Dice similarity coefficient of 0.769, a 95th percentile Hausdorff distance of 8.937, a negative Jacobian determinant rate of 0.029, and an inference time of 0.29 s. These results confirm the effectiveness and practical applicability of the proposed network in complex deformation medical image registration tasks.