Data Augmentation Using Generative Neural Networks Based on Fourier Feature Mapping

Over the past few years, Artificial Intelligence has achieved significant performance in many fields. In artificial intelligence techniques, deep neural networks have experienced rapid development recently. They have particularly excelled in image recognition, natural language processing, and speech recognition. Deep neural networks simulate how the human brain learns and recognizes objects. By providing sufficient training data, it automatically learns the hidden information in the data. However, it is common to encounter imbalanced data in real-world applications, where the quantities of data for different classes are uneven. This can make the model biased towards majority classes when making predictions. Therefore, this paper proposes two data augmentation methods by using a generative neural network based on Fourier feature mapping to generate images for data augmentation. We first The first approach primarily focuses on enhancing the diversity of image details to augment the data. The second approach primarily focuses on improving the diversity of image structures to augment the data. We also analyze the settings of the weights for the generated images. In experiments, we use a public benchmark dataset to evaluate the effectiveness of the proposed approach. The results reveal the proposed method really achieves high-quality of data augmentation.