Recently, the problem of learning with few shot has gained significant attention in the field of deep learning. The limited number of labelled samples can result in overfitting while training models, thereby increasing the difficulty level. Our proposal in this paper, DCPNet, aims to overcome the challenging classification problem using the prototype network from two aspects- feature extraction and correction of a priori feature distribution. Our approach combines a two-branch ResNet18 as the encoder and the triplet loss as the feature extraction module. We further migrate the statistical information of the a priori features to correct distribution and generate new features for input into the classifier to perform the classification. We conducted experiments on three publicly available datasets and compared our approach with classical methods, outperforming on two and achieving comparable performance with other methods on the third dataset.Our method achieves a performance improvement of about 10 percentage points over other classical methods on three datasets.