DCGAN-Based Data Augmentation for Document Classification

Document classification is a relevant task within every intelligent document processing system. With the advances in deep learning and computer vision techniques, this task has become a painless and straightforward process. However, the need for labelled data is always a hurdle to tackle before constructing, training, and validating classification models. The use of Generative modelling implies using a model to generate new samples that are similar to the training set but individually different from existing records. In this paper, we investigate using deep convolutional adversarial networks (DCGAN) to generate fake document images using existing scanned documents dataset. Moreover, we used a dataset created of generated images alongside original images to train an image classifier using a convolutional neural network (CNN) to classify documents, and we compared the overall accuracy with a model constructed on an original dataset of the same size. The constructed model performed as well as the model trained with authentic data (with an accuracy score of 90% and 91% respectively). This excellent performance can permit using DCGANs to augment existing datasets in case of a lack of labelled data while maintaining similar performance levels.