Textbook in PDF format This book comprehensively and systematically introduces the theory of generative adversarial networks and its applications in image and voice processing. This book consists of 12 chapters, of which the first four chapters present basic knowledge, including the principle of GAN, optimization objectives, training methods and evaluation indicators. The last eight chapters introduce the applications of GAN in various vertical fields, covering image generation, video generation, image translation, face image editing, image quality improvement, general image editing, anti-attack, voice signal processing and other fields. Through reading this book, readers will thoroughly understand the principles of GAN, various GAN model designs, and learn to apply GAN for most vision and voice tasks. This book is suitable for the junior researchers, students or industrial practitioners in related areas. The Chapter 1 provides a detailed introduction to unsupervised learning, supervised learning, and semi-supervised learning, including the definitions, essences, common scenarios, and frequently used models of different learning methods. Subsequently, generative models and discriminative models are respectively introduced respectively within the scope of supervised learning, covering their definitions, differences, common models, etc. Then, the concept and learning approach of unsupervised generative models are presented. In the second part of this chapter, we classify generative models into two types, explicit generative models and implicit generative models, according to the way generative models handle the probability density function. For explicit generative models, the principle of the maximum likelihood method is described in detail and is divided into two categories: tractable probability density functions and approximate methods. In the first category, FVBN series models are listed, including PixelRNN, PixelCNN, NADE, and flow models. In the second category, variational autoencoders and restricted Boltzmann machines are introduced. In the third part, the implicit generative model is introduced taking GAN as an example, and GAN is compared with other generative models. Although GAN has been widely applied in various aspects, training a GAN is not an easy task. During the training process, problems such as mode collapse, non-convergence of the loss, and blurriness of generated samples may occur. This chapter introduces the three most common problems in GAN training, namely gradient vanishing, non-convergence of the objective function, and mode collapse, and analyzes the causes of these problems. Regarding the problem of gradient vanishing, the annealing noise method is introduced. In response to the oscillation and instability of the objective function during GAN training, two methods, spectral regularization SNGAN and consistent optimization, are elaborated in detail. In addition, many GAN training techniques, such as feature matching and historical mean, are also explained. For the problem of mode collapse, from the two perspectives of the objective function and the GAN structure, some algorithms that can effectively alleviate mode collapse are introduced, and specific methods such as unrolledGAN, DRAGAN, MADGAN, VVEGAN, and the minibatch discriminator are provided. Generative Model Objective Function of GAN Training of GAN Evaluation and Visualization of GAN Image Generation Image Translation Face Image Editing Image Quality Enhancement 3D Image and Video Generation General Image Editing Adversarial Attack Speech Signal Processing