Abstract: The generative adversarial network is often used for image conversion tasks such as image coloring, semantic composition, style transfer, etc. However, the training of image generation models at this stage often depends on a large number of paired datasets, and can only achieve the conversion between two image domains. To solve the above problems, a content and style transfer based on generative adversarial network (CS-GAN) is proposed. The model maximizes the mutual information between fashion items and generated images by using a contrastive learning framework, which can ensure that content migration can be achieved without changing the structure of fashion items. Through layer to layer dynamic convolution method, style features are adaptively learned for different style images to achieve arbitrary style migration of fashion items. Content features (such as monet style and cubism) and style features (such as color and texture) of imported fashion items are integrated to achieve conversion of multiple image domains. Comparative experiments and results analysis are conducted on the open fashion data set. Comparative experiments and results analysis are carried out on the public fashion data set. Compared with other mainstream methods, this method has improved in image synthesis quality, average Inception score and FID distance evaluation indicators.

Key words: generative adversarial network, content and style transfer, feature fusion, multi-domain feature transfer, layer consistence dynamic convolution

摘要： 生成对抗网络常常被用于图像着色、语义合成、风格迁移等图像转换任务，但现阶段图像生成模型的训练往往依赖于大量配对的数据集，且只能实现两个图像域之间的转换。针对以上问题，提出了一种基于生成对抗网络的时尚内容和风格迁移模型（content and style transfer based on generative adversarial network，CS-GAN）。该模型利用对比学习框架最大化时尚单品与生成图像之间的互信息，可保证在时尚单品结构不变的前提下实现内容迁移；通过层一致性动态卷积方法，针对不同风格图像自适应地学习风格特征，实现时尚单品任意风格迁移，对输入的时尚单品进行内容特征（如颜色、纹理）和风格特征（如莫奈风、立体派）的融合，实现多个图像域的转换。在公开的时尚数据集上进行对比实验和结果分析，该方法与其他主流方法相比，在图像合成质量、Inception score和FID距离评价指标上均有所提升。

关键词: 生成对抗网络, 内容和风格迁移, 特征融合, 多域转换, 层一致性动态卷积