Conclusions

GAN is a powerful model for image generation, which has been a fundamental basis for numerous computer vision tasks. In this book, we have described several models for image generation and multi-domain image generation. For image generation, we introduced the Least Squares Generative Adversarial Networks (LSGANs), which use the least squares loss for both the generator and discriminator, instead of using the cross-entropy loss. The idea of how to overcome the vanishing gradients problem during GANs learning was explained via both intuitive examples and theoretical analysis. The results showed that LSGANs not only generate higher-quality images but also has a more stable performance than regular GANs. For multi-domain image generation, we introduced the Regularized Conditional Generative Adversarial Networks (RCGANs), which use the conditional GANs and two regularizers to force the model to encode the domain information in the conditioned domain variables. One regularizer is added to the first layer of the generator to guide the generator to decode similar high-level semantics. The other is added to the last hidden layer of the discriminator to force the discriminator to output similar losses for the corresponding images. We also introduced a method of applying RCGAN to unsupervised domain adaptation.