Thesis download address: https://arxiv.org/pdf/1610.09585.pdf
1. What problems does the article mainly solve?
In this article, we introduce a new method for image synthesis by improving Generative Adversarial Network (GAN). A variant of GAN that uses label adjustment is constructed, which results in a 128 × 128 resolution image sample that shows overall consistency. The previous image quality assessment work is expanded to provide two new analysis methods to assess the resolvability and diversity of samples from the class-conditional image synthesis model. These analyses indicate that high-resolution samples provide category information that does not exist in low-resolution samples. Among the 1000 ImageNet classes, the resolution of the resolvable 128 × 128 samples is more than twice that of manually adjusted 32 × 32 samples. In addition, 84.7% of the samples in the category are as diverse as real ImageNet data.
Applying the data set using labels to the generative confrontation network GAN can enhance the existing generative model and form two optimization ideas:
1. cGAN uses auxiliary label information to enhance the original GAN, and uses label data to train both the generator and the discriminator, so that the model has the ability to generate specific condition data.
cGAN: Conditional Generative Adversarial Nets
Article link: https://arxiv.org/pdf/1411.1784.pdf
