In order to realize the accurate registration of cross-source point clouds with different measurement scales, resolutions, and accuracy, a measurement point cloud data registration method based on multi-scale sampling is proposed. The scale slip algorithm is used to filter out the high-frequency details, retain the contour data, and combine the voxel grid neighborhood method to realize the downsampling of point cloud data. For the low-resolution point cloud data measured by macro-structured light vision, through the progressive three-dimensional point cloud upsampling algorithm based on depth learning, the contour details of structured light point clouds can be accurately restored, and the unity of scale and resolution of cross-source point clouds can be realized. Finally, the iterative nearest point method is used to register the data with scale approximate after processing, and the registration relationship is inversely applied to the registration of the original cross-source point cloud. The experimental results show that the multi-scale sampling method can improve the registration accuracy of cross-source point clouds and can be effectively used for high-performance detection of engine blades and other parts.