Partial point cloud registration stands as a pivotal pre-processing stage within computer vision applications, such as robotics, medical imaging, and autonomous driving. Previous feature-based methods rely on the quality of feature extraction to accurately estimate rigid transformations, but non-overlapping regions can easily resist this approach. Enhancing the information quality of overlapping regions and mitigating the side effects of non-overlapping regions in point cloud alignment algorithms hold considerable significance. In order to achieve this objective, we propose an unsupervised mask-based matching enhancement method for overlapping region matching. In our method, we introduce a mask generation module to obtain overlapping information while leveraging EdgeConv to acquire neighborhood information. Subsequently, we integrate this information into the point-wise matching map through the point-wise matching enhancement module, aiming to refine the estimation of point-wise matching, which leads to the generation of a more sophisticated pseudo-target point cloud through establishing more reliable correspondences. Our method demonstrates superior performance relative to existing methods, as evidenced by the experimental results on the ModelNet40, 7Scenes, and ICL-NUIM datasets.