DATA GENERATION WITH STRUCTURE ENFORCING ADVERSARIAL LEARNING

Class imbalance issues are very common among real-world datasets. Traditional oversampling approaches are interpolation based and are not well suited for image datasets. These techniques lead to class overlapping and the generation of visually unappealing minority class images. Lately, Generative Adversarial Network (GAN)-based models are used widely for oversampling of image data; however, the learning bias towards the majority classes lead to generation of majority classes in excess and minority classes in rarity. Most of the existing oversampling techniques work on data space, whereas low-dimensional latent space for oversampling is less explored. To tackle these issues, we propose a novel latent space oversampling framework called Structure Enforcing Adversarial Learning (SEAL). In the proposed architecture, the generator is trained by additionally minimizing the structure loss. This boosts the generation of synthetic samples, which retain the covariance structure of each class. The proposed model is evaluated on four image datasets and is compared with the state-of-the-art methods.