Research On Image Animation Style Transfer Based On Improved Generative Adversarial Networks

With the continuous progress of image processing technology and deep neural networks,animation technology has gradually become a new form of image creation.Transform realistic scenes and portraits into animation styles to create a brand new visual experience.Animation style conversion of images is widely used in multimedia businesses such as short video operations,internet social networking,and film and television production.Style transfer technology can process facial images through card generalization to obtain personalized animation style images.The current mainstream generation confrontation network models are prone to introduce a large number of defects that affect the appearance when migrating image styles.To solve this problem,this paper improves the image style migration method based on generating confrontation networks in the following aspects and designs experiments to verify it.Firstly,based on the AnimeGANv2 network structure,a self attention mechanism is added to the generator and discriminator to better capture long-range dependencies in images and improve the effect of image style transfer In image style transfer,the self attention mechanism can help the generator better capture the global information of the image,thereby generating more realistic images Adding a self attention mechanism to the discriminator can better distinguish the differences between images of different styles.Then,an auxiliary classifier without fully connected layers is added to the generator to learn the feature representation of the middle layer of the generator In the animation style migration of images,the auxiliary classifier can help the generator better learn the Semantic information of different animation styles and apply it to the generated images.Afterwards,different normalization methods were used to replace and compare the normalization layers in the generated adversarial network,and the best normalization method for network training was found Through comparative experiments on four normalization methods: Instance Normalization(IN),Layer Normalization(LN),Group Normalization(GN),and Switchable Normalization(SN),it can be seen that different normalization methods have different effects on the transfer of anime style,among which LN can achieve the most satisfactory effectFinally,in the model comparison experiment,we compared the proposed model with several classic image style transfer models In the ablation experiment,this article gradually removed various improvement points from the self attention mechanism,auxiliary classifier,and normalization layer replacement experiment,and conducted experimental comparisons.The experimental results show that using the improved method to transfer the style of portrait animation images can obtain images with significantly improved visual effects.The method proposed in this article can be used in animation production,game development,and has a broad application prospect.