<p>Class imbalance has always been a common and severe challenge in the field of machine learning. Due to the fact that the number of samples of the minority class is much smaller than that of the majority class, this leads to the model being overly biased towards the majority class during training, thereby reducing its ability to recognize the minority class. To alleviate this problem, the Synthetic Minority Oversampling Technique (SMOTE) has been widely adopted. However, its interpolation mechanism introduces noise samples in overlapping areas, resulting in blurred decision boundaries. To solve these problems, this paper proposes a method called Non-Boundary Minority Instance Oversampling (NB-MIO). This method aims to accurately identify and utilize minority class samples to generate new samples through a two-stage Tomek link strategy. Specifically, in the first stage, Tomek links are used to identify and remove majority class samples in the boundary area, thereby cleaning the overlapping area. In the second stage, Tomek links are re-detected, and ’safe’ minority class samples that do not form Tomek links are selected as the base instances, thereby avoiding generating new samples in the boundary area. A large number of experiments on 17 benchmark datasets with 15 methods show that NB-MIO has significant advantages. Additionally, the statistical experiment of Friedman test indicates that our method has the highest average ranking.</p>