一份簡短的風格遷移調研結果

style_transfer_survey

A survey on style_transfer from the original fantasy paper till now.

Contents:

• Papers
• Practice
• Paper Reading Notes
• Code Myself
• References

Papers

• A Neural Algorithm of Artistic Style
  • arxiv: 1508.06576
  • github: https://github.com/jcjohnson/neural-style
  • translation: https://www.jianshu.com/p/9f03b61fdeac
• Texture Networks: Feed-forward Synthesis of Textures and Stylized Images
  • arxiv: 1603.03417, also ICML 2016
  • github: author torch,
  • translation: https://www.jianshu.com/p/1187049ae1ad
• Perceptual Losses for Real-Time Style Transfer and Super-Resolution
  • arxiv: 1603.08155, also ECCV 2016
  • github: author torch, third pytorch
  • translation: https://www.jianshu.com/p/b728752a70e9
  • webpage: project page
• Incorporating long-range consistency in CNN-based texture generation
  • arxiv: 1606.01286
• Instance Normalization: The missing Ingredient for Fast Stylization
  • arxiv: 1607.08022
  • github: author torch
  • translation: https://www.jianshu.com/p/d77b6273b990
• Image Style Transfer Using Convolutional Neural Networks
  • arxiv: CVPR16
  • github: author pytorch
• A Learned Representation For Artistic Style
  • arxiv: 1610.07629, also ICLR 2017
  • github: tf
• Controlling Perceptual Factors in Neural Style Transfer
  • arxiv: 1611.07865, also CVPR 2017
  • github: author pytorch
• Fast Patch-based Style Transfer of Arbitrary Style
  • arxiv: 1612.04337
  • github: author torch
  • reference
    • http://blog.csdn.net/wyl1987527/article/details/70476044
    • http://blog.csdn.net/Hungryof/article/details/61195783
    • http://mathworld.wolfram.com/FrobeniusNorm.html
• Demystifying Neural Style Transfer
  • arxiv: 1701.01036
• Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization
  • arxiv: 1703.06868, also ICCV 2017
  • github: author torch, third pytorch
  • webpage: project
• Deep Photo Style Transfer
  • arxiv: 1703.07511
  • github: author torch, https://github.com/LouieYang/deep-photo-styletransfer-tf
  • translation: http://blog.csdn.net/cicibabe/article/details/70868746
• Neural Style Transfer: A Review
  • arxiv: 1705.04058
  • github: https://github.com/ycjing/Neural-Style-Transfer-Papers
• Universal Style Transfer via Feature Transforms
  • arxiv: 1705.08086, also NIPS 2017
  • github: author torch, third pytorch

Practice

First, pytorch has a official example fast_neural_style.

Points:

• training phase
  • content image x
  • style image s
  • pretrained model F[1,2,3,4], different middle-level feature representations on high dimension,
    (VGG16). Freezed weight parameters
  • Style Transfer Model T , a FCN model with size invariant
  • loss = weight_content * loss_content + weight_style * loss_style
• evaluating phase
  • content image x
  • trained Style Transfer Model
  • styled image y = T(x)

More Details:

• primary criterion is MSELoss.
• loss_content is criterion(F2(x), F2(y))
• GramMatrix G has nothing with the image size, is :

def gram_matrix(y):
    (b, ch, h, w) = y.size()
    features = y.view(b, ch, w * h)
    features_t = features.transpose(1, 2)
    gram = features.bmm(features_t) / (ch * h * w)
    return gram

• gm_s = [G(F1(s)), G(F2(s)), G(F3(s)), G(F4(s))]
• gm_y = [G(F1(y)), G(F2(y)), G(F3(y)), G(F4(y))]
• loss_style = sum([MSELoss(gm_s[i], gm_y[i]) for i in range(len(gm_s))])
• padding is reflection, not constant 0

Size Analysis:

• x.shape=(m1, n1, 3)
• s.shape=(m2, n2, 3)
• batch_size = b
• T downsample two times, both int(ceil(x/2)), this will bring size difference.
  For example, input image is size(3,33,33), output size is (3,36,36).
  Same as F. Saying proper.
• gm_s[i] size is (b, ch[i], ch[i])

Think About The Model:

• The VGG16 is just a representation on high dimension. It can be replaced by any other
  similar pretrained model.
• The four middle-level representations can also be chosen as other.
• Init convolution layer has a big kernel size to have a bigger receptive field.

Paper Reading Notes

• Perceptual Losses for Real-Time Style Transfer and Super-Resolution

Code Myself

• fast style transfer

References

• A Brief Summary on Neural Style Transfer

更多版本記錄可以見github oneTaken/style_transfer_survey