Multi-style scene generation method based on multi-style data set

The invention discloses a multi-style scene generation method based on a multi-style data set. Firstly, a given scene data set and a multi-style cartoon data set are preprocessed, and unification is achieved by adjusting the image resolution. Secondly, designing a network framework based on a multi-head discriminator, a multi-head style encoder and a generator; the multi-head style encoder maps multiple style codes by utilizing Gaussian noise. The input of the generator comprises a scene picture and a style code, and the output of the generator is a specific cartoonized picture. The discriminator is formed by combining a feature extraction module and a plurality of discriminating heads, and each head supports the distinguishing effect of one cartoon style and other styles. According to the method, the multi-head discriminator and the multi-head style encoder are integrated, and the characteristics of different cartoon styles are learned more comprehensively by introducing multiple style data sets. Through the