The Research of Anime Character Portrait Generation Based on Optimized Generative Adversarial Networks

Anime Character Portrait Generation is an interesting but a challenging job. At present, most of the existing methods are solved by using generative adversarial networks. However, the styles of anime character portrait are quite different and Generative Adversarial Networks is also prone to cause problems including mode collapse, which makes it hard to generate good anime character portrait samples. In this paper, DRAGAN model will be optimized and the improved version of the basic structure of Conv-BN-Relu-Pooling will be applied reasonably. The optimized DRAGAN is trained by using alternating gradient updates procedure to achieve high-quality generation samples. At the same time, through the experimental comparison with other generative adversarial networks of deep learning, it is finally verified that the optimized DRAGAN performs better than the experimental comparison group in terms of image visual quality, FID and gradient penalty terms. The mode collapse problem has also been alleviated finally.