A Data Augmentation Model Based on Variational Approach

The labeled training data are very rare in actual environment. Generating new data based on given label is one of the most commonly approaches in data augmentation. This paper proposes a new data augmentation model that can extract the deformation features between the given deformation image and the original image. The model generates similar images to the given deformation images according to the deformation feature. The model can keep the new generation images have the same probability distribution as the given deformation images. Experiments on MNIST and CIFAR-10 prove that the new deformation images can get a similar classification accuracy with the given deformation images, which proves that the new sample is effective.