Research on the Generation of Creative Animation Driven by Deep Learning Model

Abstract

It is a very interesting and practical task to transform real-world images such as portraits or scenery into creative animation images. Since this concept was put forward, it has aroused extensive research interest in the field of computer vision. The generative adversarial networks (GAN) model is widely used in this field. Depth convolution GAN (DCGAN) and Wasserstein GAN (WGAN) improve the original GAN, but there are still problems existing in creative animation generation such as model collapse. To solve these problems, the Wasserstein distance is introduced to replace the JS divergence in the GAN model to measure the gap between the sample distribution generated by the generator and the real distribution, and the loss function is improved. In order to achieve a better animation generation effect, the training of the model is further optimized through the adjustment of the network model structure and the setting of parameters. Through the comparison with DCGAN and WGAN models in the animation data set and CelebA data set and the quantitative analysis and comparison of the generation effects of different models, the effectiveness and generalization of the improved GAN model are verified.