Deep learning has been widely used ever since convolutional neural networks (CNN) have shown great improvements in the field of computer vision. Developments in deep learning technology have mainly focused on the discriminative model; however recently, there has been growing interest in the generative model. This study proposes a new model that can learn disentangled representations and generate high quality images. The model concatenates the latent code to the noise in the training process and maximizes mutual information between the latent code and the generated image, as shown in InfoGAN, so that the latent code is related to the image. Here, the concept of balancing between discriminator and generator, which was introduced in BEGAN, is adapted to create better quality images under high-resolution conditions.
展开▼