keras神经风格迁移_深度神经风格迁移
介紹(由于近期準(zhǔn)備校招,博客暫時(shí)不更新)
神經(jīng)風(fēng)格遷移是我研一研二時(shí)期主要的研究方向,而從最初的風(fēng)格遷移出現(xiàn)已經(jīng)有較長(zhǎng)一段時(shí)間了。之所以現(xiàn)在寫這個(gè)博客,第一是因?yàn)槲业漠厴I(yè)論文定的方向是風(fēng)格遷移+情感分析;第二是借這篇博客以及之后的學(xué)習(xí),能對(duì)深度學(xué)習(xí)更進(jìn)一步的理解!
目錄
什么是風(fēng)格遷移?最經(jīng)典的Gatys(paper,code)中給出了風(fēng)格遷移的流程圖,視為最原始的方法。借此圖說(shuō)明風(fēng)格遷移的基本原理。
α 圖定義為style image, p 圖定義為content image,
損失通過(guò)VGG-16的前四層來(lái)表示,層次越高,內(nèi)容越抽象。這里列出幾個(gè)符號(hào)定義。
將內(nèi)容圖像輸入卷積網(wǎng)絡(luò)中提取圖像內(nèi)容,由公式
,計(jì)算內(nèi)容損失。對(duì)以上公式求導(dǎo)
,使用反向傳輸,使得生成圖像在內(nèi)容上接近原輸入內(nèi)容圖像。將風(fēng)格圖像輸入到同一個(gè)網(wǎng)絡(luò)中提取它的風(fēng)格信息,風(fēng)格提取的符號(hào)定義為
計(jì)算風(fēng)格圖像的loss
單獨(dú)某層的損失函數(shù)
各層綜合的損失函數(shù)
求偏導(dǎo)
,使得生成圖像在風(fēng)格上接近原輸入風(fēng)格圖像。7.風(fēng)格損失
8.Gatys-Image-Style-Transfer中給出的流程圖。X是白噪聲圖像。同時(shí)將三張圖片輸入到同一網(wǎng)絡(luò)中,對(duì)內(nèi)容圖像和風(fēng)格圖像求特征,對(duì)白噪聲X求導(dǎo)。
當(dāng)前不同框架下的風(fēng)格遷移
幾年前,Gatys等人的風(fēng)格遷移[paper],[code]在學(xué)術(shù)界引起了不錯(cuò)的反響,并催生了后續(xù)很多研究成果。雖然在Gatys之前已經(jīng)有學(xué)者做遷移方面的研究,但我把這篇paper看作是first style transfer paper。
1.基于圖像優(yōu)化的Slow Transfer[A Neural Algorithm of Artistic Style ][paper]
[Demystifying Neural Style Transfer][paper](Theoretical Explanation) (IJCAI 2017)
[Stable and Controllable Neural Texture Synthesis and Style Transfer Using Histogram Losses][paper]
[Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis][paper](CVPR 2016)
2.基于模型優(yōu)化的Fast Transfer
2.1Per-Style-Per-Model-Methods[Perceptual Losses for Real-Time Style Transfer and Super-Resolution][paper] (ECCV 2016)
[Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks] [Paper] (ECCV 2016)
[Texture Networks: Feed-forward Synthesis of Textures and Stylized Images] [Paper] (ICML 2016)
[Improved Texture Networks: Maximizing Quality and Diversity in Feed-forward Stylization and Texture Synthesis] [Paper] (CVPR 2017)
[Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks] [Paper] (ECCV 2016)
2.2Multiple-Style-Per-Model-Methods[A Learned Representation for Artistic Style] [Paper] (ICLR 2017)
[Multi-style Generative Network for Real-time Transfer] [Paper]
[StyleBank: An Explicit Representation for Neural Image Style Transfer] [Paper] (CVPR 2017)
[Diversified Texture Synthesis With Feed-Forward Networks] [Paper] (CVPR 2017)
2.3Arbitrary-Style-Per-Model-Methods[Fast Patch-based Style Transfer of Arbitrary Style] [Paper]
[Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization] [Paper] (ICCV 2017)
如何評(píng)價(jià)風(fēng)格遷移的實(shí)驗(yàn)效果
總結(jié)
以上是生活随笔為你收集整理的keras神经风格迁移_深度神经风格迁移的全部?jī)?nèi)容,希望文章能夠幫你解決所遇到的問(wèn)題。
- 上一篇: 南邮CTF web题目总结
- 下一篇: python3操作sqite3删除数据表