Data augmentation for handwritten digit recognition using generative adversarial networks

Supervised learning techniques require labeled examples that can be time consuming to obtain. In particular, deep learning approaches, where all the feature extraction stages are learned within the artificial neural network, require a large number of labeled examples to train the model. Various data augmentation techniques can be performed to overcome this issue by taking advantage of known variations that have no impact on the label of an example. Typical solutions in computer vision and document analysis and recognition are based on geometric transformations (e.g. shift and rotation) and random elastic deformations of the original training examples. In this paper, we consider Generative Adversarial Networks (GAN), a technique that does not require prior knowledge of the possible variabilities that exist across examples to create novel artificial examples. In the case of a training dataset with a low number of labeled examples, which are described in a high dimensional space, the classifier may generalize poorly. Therefore, we aim at enriching databases of images or signals for improving the classifier performance by designing a GAN for creating artificial images. While adding more images through a GAN can help, the extent to which it will help is unknown, and it may degrade the performance if too many artificial images are added. The approach is tested on four datasets on handwritten digits (Latin, Bangla, Devanagri, and Oriya). The accuracy for each dataset shows that the addition of GAN generated images in the training dataset provides an improvement of the accuracy. However, the results suggest that the addition of too many GAN generated images deteriorates the performance.