In the typical pipeline of three-dimensional local feature-based registrations, the Local Reference Frame (LRF) is significant in achieving the pose normalization of local descriptor, thereby enhancing the accuracy of feature correspondence. However, existing LRFs often struggle to maintain high repeatability under strong Gaussian noise and high mesh decimation interference. To address this challenge, this paper presents a novel LRF. It constructs the Z-axis using the smallest feature vector of a smaller-scale surface and constructs the X-axis leveraging the largest feature vector of a larger scale. Simultaneously, it addresses noise and uneven point density interference using weight terms. Experimental results substantiate that the presented LRF effectively maintains high repeatability under strong noise and high decimating rates, thereby improving feature matching accuracy and benefiting tasks such as point cloud registration.