第二篇论文解读 APDrawingGAN
第二篇论文解读 APDrawingGAN
Conclusion
- 这篇论文做出了什么成果?
- 提出了一个层次化的GAN模型,可以有效地将人脸照片生成高质量、富有表现力的艺术肖像线条画。不仅如此,我们的方法对黑白线条分明的复杂发型绘制有更好的效果。
- 为了学习不同面部区域的不同绘制风格,我们的模型将GAN的渲染输出分为不同层次,每个层次被独立的损失项控制。我们提出了一个针对艺术肖像画的损失函数,它包含四个损失项:对抗损失、像素级损失、一种新的距离变换(DT)损失(用于学习艺术肖像画中的线条笔画)和一个局部变换损失(用于引导局部网络保持面部特征)
- 构建预训练数据集和正式训练集
- 相关的工作有哪些?作者的观点/成果和别人的突出在哪里?
- Style transfer using neural networks
- In addition to aforementioned limitations for APDrawing style transfer, most existing methods require the style imageto be close to the content image.
- Non-photorealistic rendering of portraits
- However, all these methods use similar texture synthesis approaches that make them unsuitable for the APDrawing style
- GAN-based image synthesis
- Neither Pix2Pix nor CycleGAN works well for APDrawing styles and often generates blurry or messy results due to the five challenges summarized in Sec. 1 for APDrawings.
- 上述滴问题,本文中提出的APDrawing GAN都可以在一定程度上解决
- Style transfer using neural networks
- 这篇论文是如何论证的?作者使用了哪些论据来支持观点或者成果。
- 通过比较现在已有的方法-如 Gatys, CNNMRF, Deep Image Analogy , Pix2Pix , CycleGAN and Headshot Portrait.
- 用户体验
论文大体介绍
-
首先本文提出艺术肖像线条画(Artistic Portrait Drawings,简称APDrawings)和已有工作研究的油画肖像的风格有很大的不同。它主要有5个特点:
-
得出结论,即使是顶尖的方法也难以产生好的艺术肖像画结果。
-
-
提出自己的APDrawing GAN模型,优点总体如下
-
hierarchical GAN
- This allows dedicated drawing strategies to be learned for different facial features
-
a novel loss
- Since artists’ drawings may not have lines perfectly aligned with image features, we develop
a novel loss to measure similarity between generated and artists’ drawings based on distance transforms, leading to improved strokes in portrait drawing.
- Since artists’ drawings may not have lines perfectly aligned with image features, we develop
-
hierarchical GAN
-
讲解模型
- 思维导图如下(内容较多,麻烦请放大观看)
- 此外还有Hierarchical generator/Hierarchical Discriminator的图解(来源自论文)
- 思维导图如下(内容较多,麻烦请放大观看)
-
如何训练
- 预训练
- 正式训练
-
Experiments
- Ablation study in APDrawingGAN
- Comparison with state-of-the-art