A prototype correction method based on Gaussian distribution is proposed to address the issues of class prototypes being prone to bias and poor network generalization that arise in metric-based few-shot learning methods due to the scarcity of support samples. The proposed method first obtains the class prototypes based on the prototypical network, and perform nearest neighbor matching on the query samples through the class prototypes to get the pseudo labels of the query samples. Then, the Gaussian distribution information of the pseudo labeled sample features is acquired, namely mean and variance. Finally, enough samples are generated by sampling from these distributions to expand the support set, thus obtain more accurate class prototypes and improve classification performance. At the same time, orthogonal constraints are introduced into the existing feature extraction network to improve the generalization of the model. The few-shot classification experiments and further ablation experiments are performed on common point cloud datasets. On the ModelNet40 and ModelNet40-C datasets, the average classification accuracy of the proposed method is comparable to the existing method. On the noisy ScanObjectNN and ScanObjectNN-PB datasets, the average classification accuracy is better than the existing method by 1.36%. The further ablation experiments verify the effectiveness of prototype correction and network parameter constraints. The proposed method can effectively alleviate the overfitting problem in few-shot point cloud classification and has strong robustness against perturbed data.