Image Data Augmentation from Small Training Datasets Using Generative Adversarial Networks (GANs)

The scarcity of labelled data is a serious problem since deep models generally require a large amount of training data to achieve desired performance. Data augmentation is widely adopted to enhance the diversity of original datasets and further improve the performance of deep learning models. Learning-based methods, compared to traditional techniques, are specialized in feature extraction, which enhances the effectiveness of data augmentation. Generative adversarial networks (GANs), one of the learning-based generative models, have made remarkable advances in data synthesis. However, GANs still face many challenges in generating high-quality augmented images from small datasets because learning-based generative methods are difficult to create reliable outcomes without sufficient training data. This difficulty deteriorates the data augmentation applications using learning-based methods. In this thesis, to tackle the problem of labelled data scarcity and the training difficulty of augmenting image data from small datasets, three novel GAN models suitable for training with a small number of training samples have been proposed based on three different mapping relationships between the input and output images, including one-to-many mapping, one-to-one mapping, and many-to-many mapping. The proposed GANs employ limited training data, such as a small number of images and limited conditional features, and the synthetic images generated by the proposed GANs are expected to generate images of not only high generative quality but also desirable data diversity. To evaluate the effectiveness of the augmented images generated by the proposed models, inception distances and human perception methods are adopted. Additionally, different image classification tasks were carried out and accuracies from using the original datasets and the augmented datasets were compared. Experimental results illustrate the image classification performance based on convolutional neural networks, i.e., AlexNet, GoogLeNet, ResNet and VGGNet, is comprehensively enhanced, and the scale of improvement is significant when a small number of training samples are involved.