Accurate class distribution estimation is expected to solve the problem of the poor generalization ability that exists in few-shot learning models due to data shortages. However, the reliability of class distributions estimates based on limited samples and knowledge is questionable, especially for similar classes. We find that the distribution calibration method is inaccurate in estimating similar classes due to limited knowledge being reused through double-validation experiments. To address this issue, we propose a novel class center estimation (NC2E) method, which consists of a two-stage center estimation (TCE) algorithm and a class centroid estimation (CCE) algorithm. The class centers estimated by TCE in two stages are closer to the truth, and its superiority is demonstrated by error theory. CCE searches for the centroid of the base class iteratively and is used as the basis for the novel class calibration. Sufficient simulation samples are generated based on the estimated class distribution to augment the training data. The experimental results show that, compared with the distribution calibration method, the proposed method achieves an approximately 1% performance improvement on the miniImageNet and CUB datasets; an approximately 1.45% performance improvement for similar class classification; and an approximately 6.06% performance improvement for non-similar class classification.