Neural Artists

项目介绍

利用基于条件生成式对抗网络的解决方案为用户提供照片和视频等影音素材的艺术风格转换服务。用户可通过网页和微信小程序等方式进行交互：用户上传图片或视频文件，选择艺术滤镜或自定义风格，经后端处理渲染素材，即可显示预览效果和发布分享。

Neural Artists converts users' multimedia materials into customized artistic styles based on conditional generative adversarial networks (CGANs). The service is provided through a web console as well as a WeChat mini-program. Users only need to upload their raw files and apply the art filters provided or a custom style. Our backend servers will process the image by pixels and render the video frame by frame.

技术原理

原始方案

用户选择的目标风格通过卷积神经网络VGG提取Neural Style[1]。 用户视频（照片）提取关键帧，逐层渲染并重构目标图像。 综合关键帧信息，输出目标视频（照片）。

进阶方案

通过Fast Style Transfer[2]框架和CGAN(Conditional Generative Adversarial Networks)实现像素层面的风格转换。通过调整损失函数和正则项平衡视频内容的协同性和风格的独特性，构建端到端的训练框架。

Reference

[1] Neural Style https://github.com/jcjohnson/neural-style
[2] Fast Style Transfer in TensorFlow https://github.com/lengstrom/fast-style-transfer

Built With

• python
• tensorflow