Generative Adversarial Network (GAN) was proposed in 2014 by Goodfellow et al. It can generate very realistic synthetic images, making the generated images and the real images almost indistinguishable in statistics.
An intuitive understanding of generative adversarial network is that if a forger forges some cat pictures, at first the forger is not familiar with how to draw the cat, he will hand the real cat pictures to the connoisseur together with his forgery works, the connoisseur will evaluate the authenticity of each work, and tell the forger the difference between the works and the real pictures, the forger update again based on this and repeat the above steps. After continuous circulation, the forger can copy very realistic cat pictures, which are more and more difficult for the connoisseur to distinguish. The forger and connoisseur constantly “fight” against each other, thus “generating” a batch of very real fake cat pictures.
Similar to the above example, the generative adversarial network consists of two parts: generator network and discriminator network. The generator network takes a random vector (random point in potential space) as input and decodes it into an image. The discriminator network uses a picture as input to predict whether it comes from training set or generator network. With the continuous training, the generator network can continuously cheat the discriminator to generate more and more realistic pictures. At the same time, the discriminator network is also constantly high-capacity, which sets a higher standard for the generator network.
Now I’m trying to use GAN to generate leaf images and provide data sets for further research.
Some generated images