Sample extraction and expansion method with feature reconstruction and deformation information

Neural networks often need a large number of data to complete effective training. In the low-data regime, networks perform poor in training effect and generalization ability. Recently,most few-shot learning methods based on the data expansion generate data by utilizing generative adversarial idea,or fulfill data augmentation by directly adopting difference information between similar datasets. The data after expansion will still be lack of important data features without considering whether the features of few-shot are complete before expansion. Nor does it examine whether the adoption of difference information is reasonable, which will generate wrong samples. Therefore, this paper puts forward an adversarial data augmentation model based on feature reconstruction and deformation information. Firstly, it proposes sample extraction method based on feature reconstruction, which is used to improve the feature loss of few-shot, and it adopts feature reconstruction to extract typical few-shot for sample expansion. Moreover, it puts forward the sample expansion method based on deformation information, and it adopts deformation information of different clusters under the same class to fulfill the data expansion. The above mentioned methods are applied to the character datasets and some popular few-shot datasets. The typical few-shot after reconstruction and the dataset after expansion have good effects. Furthermore, the experiment results demonstrate the state-of-the-art performance and effectiveness of the proposed methods.