Robust point features are essential when registering point cloud scenes with numerous instances. To enhance the point features, we propose KPConvFormer module. It leverages the advantages of attention mechanisms to focus on important features, considers the feature and position differences among points in point convolution simultaneously, and pre-weights the neighborhood points in the convolution region. The pre-weight process filters out irrelevant points from other instances near the boundaries and noisy points within the convolution region, correcting the point convolution factors in each neighborhood to help aggregate more accurate point features. Addressing the incorrect registration caused by the similar structure of point clouds, we designed a Shareped-Linear-Self-Attention module. It learns a sharp global prior, efficiently capturing fine-grained global structural information. This module distinguishes similarity structures in the point clouds to be registered from a larger receptive field, providing a global prior for subsequent convolution operations. Compared to existing state-of-the-art methods, our approach achieves superior performance on most registration metrics across the 3DMatch, 3DLoMatch, and KITTI datasets.