Abstract:  With the advancement of digital dentistry and 3D oral medicine, high-precision 3D dental models have been widely used in orthodontic planning, prosthetic design, and surgical simulation. However, dental models often consist of tens of thousands of triangular meshes, and such high-density meshes pose significant burdens on data storage, transmission, and real-time interaction. Existing geometric-driven and attribute-driven mesh simplification methods perform well in general scenarios, but they still face the challenges of feature loss and boundary breakage when dealing with the pronounced regional geometric heterogeneity and clinical feature centralization in dental models. This paper proposes a mesh simplification method for dental models that integrates region-adaptive partitioning and cross-region feature preservation. The method first partitions the dental model based on normal similarity and breadth-first search (BFS) of dynamic average normals, distinguishing flat surfaces from complex occlusal surfaces, and formulates differentiated simplification strategies for different regions. Subsequently, a cross-region feature point detection and preservation mechanism is employed, giving higher preservation weights to vertices located at the boundaries between multiple regions, thus retaining key structures such as cusps, grooves, and occlusal edges under high simplification rates. Experimental results on real dental models show that the proposed method achieves low geometric deviation and high normal similarity at simplification rates of 80%, 50%, and 10%. Compared with mainstream methods like QEM, Meshlab, LPM, and ICE, the proposed method performs better in terms of Hausdorff distance, average distance, and Chamfer distance, and it effectively preserves clinically relevant geometric features even under high simplification rates. This method provides strong support for orthodontic simulation, prosthetic design, and real-time interactive applications in digital dentistry.