%A ZHANG Yang, MA Xiao-hu %T Anime Character Portrait Generation Algorithm Based on Improved Generative Adversarial Networks %0 Journal Article %D 2021 %J Computer Science %R 10.11896/jsjkx.191100092 %P 182-189 %V 48 %N 1 %U {https://www.jsjkx.com/CN/abstract/article_19686.shtml} %8 2021-01-15 %X In order to solve the problems of poor diversity,generation by class and detail control in existed method,we present an improved model named LMV-ACGAN.It is based on ACGAN and involved with mutual information and multiscale discrimination.Our model includes a feature combined generator,a multiscale discriminator and three fully connected nets for real-fake judging,classifying and latent label restoration.As a semi-supervised generative model,except class label,we also use a group of continuous latent label to enhance the constraint of the generator.Moreover,in our algorithms,pooling layers in VGG-NET are replaced by stride convolutions.Then the discriminator uses the multiscale information of the image to feature fusion.Finally,we improve the tail-end structure of the model and the rules of parameters update so as to reduce the influence between classification,real-fake judgement and latent label restoration as far as possible.Our experiment shows that the proposed method effectively solve the problem of mode collapse on our dataset,meanwhile compared with origin ACGAN,our method increases the success rate and accuracy of generating specified class image.For the image which is generated poorly or classified incorrectly by ACGAN,our method can achieve the goal.In addition,our model enable people to modify the continuous latent label to realize image editing such as changing the face orientation.