Abstract: A dual-branch, multi-level feature fusion network (DMFNet) was proposed to address point cloud registration under partial overlap, occlusion, and noise. In the encoder, a rotation branch and a translation branch were configured in parallel. Self-attention fusion and cross-attention fusion modules were inserted at shallow, middle, and deep layers so that multi-scale interaction between source and reference point clouds was achieved. A rotation–translation feature fusion module was designed to strengthen pose estimation. A lightweight Set Transformer was adopted as the regressor; multi-layer induced-attention blocks and an attention-pooling module were used to directly regress the quaternion and the translation vector. DMFNet does not rely on overlap-region detection or explicit mask estimation, and it shows strong adaptability and generalization. Comparative experiments with six registration methods were conducted on ModelNet40, and a generalization study was carried out on the Stanford 3D Scanning dataset. The results show that, under noisy conditions on the ModelNet40 dataset, DMFNet reduces RMSE(t) and Error(R) by 21.32% and 14.47% compared with MAC, and demonstrates superior robustness and registration accuracy.