A study of generative adversarial networks and possible extensions of GANs
MetadataShow full item record
The goal of our research is to explore the power of generative adversarial networks (GANs). We take a review of deep learning and many extended versions of GANs. We implement and empirically evaluate Deep Convolutional GANs. As a contribution, we propose an extension of GANs called Extractor GANs (EGANs). It contains a feature extractor to get extractable attributes of generated images. We first use the extractor to get some low-level features, such as brightness and sharpness. EGANs can determine the brightness and sharpness of the generated images with the help of the extractable attributes. In order to control more complicated features of face images, we consider to use convolutional neural networks (CNN) as the feature extractor. We use two ways, pretraining and without pre-training, to train CNN so we can have a well-performed feature extractor. The results show that EGANs perform well in controlling some high-level features of face images, such as gender and openness of mouth.