ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

稳定的图片神经网络纹理合成 – 译学馆
未登陆,请登陆后再发表信息
最新评论 (0)
播放视频

稳定的图片神经网络纹理合成

Stable Neural Style Transfer | Two Minute Papers

亲爱的学霸们大家好 这里是Károly Zsolnai-Fehér带来的两分钟论文
Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér.
神经网络风格迁移是一项超棒的技术 当我们输入两幅照片 它能够
Neural style transfer is an incredible technique where we have two input photographs, and the
输出一幅融合这两幅照片的图像 准确地说 这幅合成图像是输入图片其中一幅图的内容和另一幅图
output would be a combination of these two, namely, the content of one and the artistic
的风格相融合的结果
style of the other fused together.
当这个方向的第一篇论文出现时 关于它的新闻震惊了整个世界 并且有很多
When the first paper appeared on this topic, the news took the world by storm, and lots
臆测性的言论开始出现 讨论这种技术可以用在哪些方面和它如何能够改变
of speculative discussions emerged as to what this could be used for and how it would change
数字艺术和视频游戏工业
digital arts and the video game industry.
玩玩这些算法是非常有意思的 同时我们还见证了最近疯狂衍生的
It is great fun to use these algorithms and we have also witnessed a recent proliferation
能够实现这种功能的手机应用 这听起来太酷了 有两个原因
of phone apps that are able to accomplish this, which is super cool because of two reasons:
首先从发表的论文到应用到工业领域的时间
one, the amount of time to go from a published research paper to industry-wide application
从来没有这么短过 其次这个课题的第一篇工作需要一台强力的电脑来支持
has never been so small, and, two, the first work required a powerful computer to accomplish
并且需要好几分钟的高强度运算 而现在 仅仅不到两年的时间
this, and took several minutes of strenuous computation, and now, less than two years
仅仅用你口袋里的手机就可以实时地生成
later, it’s right in your pocket and can be done instantly.
这些指数性增长的科研进展真是令人震惊
Talk about exponential progress in science and research, absolutely amazing.
而现在 当你还沉醉于这些美轮美奂的效果中的时候 我们来谈谈
And now, while we feast our eyes upon these beautiful results, let’s talk about the selling
这项对原始算法的扩展有什么具体的卖点
points of this extension of the original technique.
这篇论文文包括非常好的关于现有的风格迁移算法的弱点的
The paper contains a nice formal explanation of the weak points of the existing style transfer
理论解释
algorithms.
这些理论解释的直观理解就是传统的基于神经网络的算法从神经元
The intuition behind the explanation is that the neural networks think in terms of neuron
活动的角度思考 而这种算法的结果和原图风格中的色彩强度不成比例
activations, which may not be proportional to the color intensities in the source image
所以他们的结果经常变得不一致或者和我们期待的不一样
styles, therefore their behavior often becomes inconsistent or different than expected.
而这篇文章的作者提出从直方图的角度思考 这意味着输出图像应该基于
The authors propose thinking in terms of histograms, which means that the output image should rely
和原图的统计相似性
on statistical similarities with the source images.
我们可以看到 这些新结果看起来真的很棒 即使是和原方法对比
And as we can see, the results look outstanding even when compared to the original method.
同时十分重要的是这项技术还更加具有艺术感
It is also important to point out that this proposed technique is also more art directable,
你一定要看看这篇论文里的细节
make sure to have a look at the paper for more details on that.
和往常一样 我已经把论文地址附在视频介绍里了
As always, I’ve put a link in the video description.
这项扩展还可以被用在纹理合成方面 这意味着我们可以输入一小块
This extension is also capable of texture synthesis, which means that we give it a small
图像 用来表示某种可重复的纹理 之后这种算法会尝试把这一小块纹理无限扩展 并且看起来
image patch that shows some sort of repetition, and it tries to continue it indefinitely in
完全无缝
a way that seems completely seamless.
不过 我们必须注意图形学领域的一个事实
However, we have to be acutely aware of the fact that in the computer graphics community,
就是纹理合成已经成为了图形学里一个独立的子领域 已经包含至少上百篇文章了
texture synthesis is considered a subfield of its own with hundreds of papers, and one
要想在这个领域发表文章 相对于现在的最新进展 作者必须有非常确定的清晰的卖点
has to be extremely sure to have a clear cut selling point over the state of the art.
在座的学霸们如果有对这个方向感兴趣的 我已经把一篇综述性文章放到
For the more interested Fellow Scholars out there, I’ve put a survey paper on this in
视频介绍里了 你们一定要看看
the video description, make sure to have a look!
谢谢大家的收看和支持 我们下次再见
Thanks for watching and for your generous support, and I’ll see you next time!

发表评论

译制信息
视频概述

这次我们介绍了一篇arxiv上的文章“基于直方图损失的稳定可控的纹理合成和风格迁移”

听录译者

收集自网络

翻译译者

GraphiCon-origamidance

审核员

知易行难

视频来源

https://www.youtube.com/watch?v=8u3Hkbev2Gg

相关推荐