Fast and High-resolution Image Generation Based on Improved DCGAN

Abstract

Among generative models, Generative Adversarial Network (GAN) has been sought after by researchers since its proposal. The researching fields of image generation, style transformation, data augmentation, super-resolution, image restoration, and image transformation have shined because of GAN. Deep Convolution Generative Adversarial Network (DCGAN), as an early neural network to improve GAN, solved the problem of unstableness during training. It can be easily scaled to deal with larger datasets and more sophisticated tasks, so various image generation and manipulation tasks can be tackled by this powerful tool. Nevertheless, it still has certain problems. This research investigates the hyperparameters, label smoothing and improved model’s effect on the quality and speed of image generation, and finally selects the appropriate hyperparameters and label smoothing to cooperate with the improved model to quickly generate clearer images with DCGAN in the case of few samples and few number of trainings. This work can bring some ideas for saving computational resources and data for training.