Training Dataset Extension Through Multiclass Generative Adversarial Networks and K-nearest Neighbor Classifier

Abstract

The performance of deep learning architectures, where all the feature extraction stages are learned within the artificial neural network, requires a large number of labeled examples to model the variability of the different possible inputs. This issue is also present in other classification tasks with a large number of features where the number of examples can limit the number of possible input features, involving the creation of handcraft feature sets. Typical solutions in computer vision and document analysis and recognition increase the size of the database with additional geometric transformations (e.g. shift and rotation) and random elastic deformations of the original training examples. In this paper, we propose to evaluate the impact of additional images created through generative adversarial networks (GANs), which are deep neural network architectures. We study the addition of images created through a multiclass GAN in different databases of handwritten numerals from different scripts (Latin, Devanagari, and Oriya). The contributions of this paper are related to the use of multiclass GANs to extend the size of the training database after filtering the images with a k-nearest neighbor classifier where the k nearest neighbors must all agree on the decision to validate a GAN generated image. The accuracy is evaluated with the original training dataset, the GAN generated images, and the combination of the original training images and the GAN generated images. The results support the conclusion that GAN generated images through a multiclass paradigm can provide a robust and fully data driven solution for enlarging the size of the training database for improving the accuracy on the test dataset.