图像修复：人类和 AI 的对决 #6684
Merged
图像修复：人类和 AI 的对决 #6684
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
图像修复：人类和 AI 的对决
Merged
|
校对认领 |
|
@lsvih 好的呢 🍺 |
|
校对认领 @Glowin |
|
@Amberlin1970 妥妥哒 🍻 |
lsvih
reviewed
Feb 9, 2020
|
|
||
| Image inpainting is the process of reconstructing missing parts of an image so that observers are unable to tell that these regions have undergone restoration. This technique is often used to remove unwanted objects from an image or to restore damaged portions of old photos. The figures below show example image-inpainting results. | ||
| 图像修复是对一幅图像丢失部分的重构过程，使得观察者察觉不到这些区域曾被修复。 这种技术通常用于移除图像中不想要的物体，或者是修复老照片上损坏的部分. 下面的图片展示了图像修复结果的例子。 |
There was a problem hiding this comment.
Suggested change
| 图像修复是对一幅图像丢失部分的重构过程，使得观察者察觉不到这些区域曾被修复。 这种技术通常用于移除图像中不想要的物体，或者是修复老照片上损坏的部分. 下面的图片展示了图像修复结果的例子。 | |
| 图像修复是对一幅图像丢失部分的重构过程，使得观察者察觉不到这些区域曾被修复。这种技术通常用于移除图像中不想要的物体，或者是修复老照片上损坏的部分。下面的图片展示了图像修复结果的例子。 |
|
|
||
| Image inpainting is an ancient art that originally required human artists to do the work by hand. But today, researchers have proposed numerous automatic inpainting methods. In addition to the image, most of these methods also require as input a mask showing the regions that require inpainting. Here, we compare nine automatic inpainting methods with results from professional artists. | ||
| 图像修复是一门古老的艺术，最初需要人类艺术家手工作业。但如今，研究人员提出了许多自动修复方法。除了图像本身，大多数方法还需要展示修复区域的遮罩（mask）作为输入。在这里，我们将对九个自动修复方法和专业艺术家进行比较。 |
There was a problem hiding this comment.
除了图像本身，大多数方法还需要展示修复区域的遮罩（mask）作为输入。
->
大多数自动修复方法除了图像本身外，还需要输入一个遮罩（mask）来表示需要修复的区域。
这样会不会通顺一点
|
|
||
| .)](https://cdn-images-1.medium.com/max/2152/1*EOuFiCNYdNde05bi9UmB8A.jpeg) | ||
| .)](https://cdn-images-1.medium.com/max/2152/1*EOuFiCNYdNde05bi9UmB8A.jpeg) |
There was a problem hiding this comment.
Suggested change
| .)](https://cdn-images-1.medium.com/max/2152/1*EOuFiCNYdNde05bi9UmB8A.jpeg) | |
| ）](https://cdn-images-1.medium.com/max/2152/1*EOuFiCNYdNde05bi9UmB8A.jpeg) |
|
|
||
| )](https://cdn-images-1.medium.com/max/2412/1*_Ldd9jY-9xS2OEE6Z8FTfw.jpeg) | ||
| )](https://cdn-images-1.medium.com/max/2412/1*_Ldd9jY-9xS2OEE6Z8FTfw.jpeg) |
There was a problem hiding this comment.
Suggested change
| )](https://cdn-images-1.medium.com/max/2412/1*_Ldd9jY-9xS2OEE6Z8FTfw.jpeg) | |
| ）](https://cdn-images-1.medium.com/max/2412/1*_Ldd9jY-9xS2OEE6Z8FTfw.jpeg) |
|
|
||
| We used a private, unpublished photo collection to ensure that the artists in our comparison had no access to the original images. Although irregular masks are typical in real-world inpainting, we stuck with square masks at the center of the image, since they’re the only type that some DNN methods in our comparison allow. | ||
| 我们使用一个私有，未公开的照片集来保证参与对比的艺术家们没有接触过原始图片。尽管不规则的遮罩是现实世界图像修复的典型特征，但我们不得不在图像中心使用正方形的遮罩，因为它们是我们对比中一些 DNN 方法所唯一允许的类型。 |
There was a problem hiding this comment.
尽管不规则的遮罩是现实世界图像修复的典型特征，但我们不得不在图像中心使用正方形的遮罩，因为它们是我们对比中一些 DNN 方法所唯一允许的类型。
->
虽然不规则的遮罩是现实世界图像修复的典型特征，但我们只能在图像中心使用正方形的遮罩，因为它们是我们对比实验中一些 DNN 方法所唯一允许的遮罩类型。
这样可能通顺一点
| @@ -42,70 +42,70 @@ We applied to our test data set six inpainting methods based on neural networks: | |||
| 5. Generative Image Inpainting With Contextual Attention ([Yu et al., 2018](https://arxiv.org/abs/1801.07892)) — this method appears twice in our results because we tested two versions, each trained on a different data set (ImageNet and Places2) | |||
| 6. Image Inpainting for Irregular Holes Using Partial Convolutions ([Liu et al., 2018](https://arxiv.org/abs/1804.07723)) | |||
|
|
|||
| As a baseline, we tested three inpainting methods proposed before the explosion of interest in deep learning: | |||
| 我们测试了三个在人们对深度学习的兴趣爆发前的提出修复方法（非神经网络方法）作为基准线： | |||
|
|
||
| We hired three professional artists who do photo retouching and restoration and asked each of them to inpaint three images randomly selected from our data set. To encourage them to produce the best possible results, we also told each artist that if his or her works outranked the competitors, we would add a 50% bonus to the honorarium. Although we imposed no strict time limit, the artists all completed their assignments in about 90 minutes. | ||
| 我们雇佣了三个从事图像后期调整和修复的专业艺术家，让他们修复从我们的数据库中随机选取的图片。为了激励他们得到尽可能好的结果。我们跟他们说，如果他或她的作品比竞争对手好，我们会给酬金增加 50% 作为奖励。尽管我们没有给定严格的时间限制，所有艺术家都在 90 分钟左右的时间内完成了任务。 |
There was a problem hiding this comment.
所有艺术家都在 90 分钟左右的时间内完成了任务。
->
但所有艺术家都在 90 分钟左右的时间内完成了任务。
|
|
||
| We compared the inpainting results from the three professional artists and the results from the automatic inpainting methods against the original, undistorted images (i.e., ground truth) using the [Subjectify.us](http://www.subjectify.us) platform. This platform presented the results to study participants in a pairwise fashion, asking them to choose from each pair the image with the best visual quality. To ensure that participants make thoughtful selections, the platform also conducts verification by asking them to compare the ground truth image and the result of Exemplar-Based Image Inpainting. It discarded all answers from respondents who failed to correctly answer one or both of the verification questions. In total, the platform collected 6,945 pairwise judgments from 215 participants. | ||
| 我们使用[Subjectify.us](http://www.subjectify.us)平台将三个专业艺术家和自动图像修复方法的结果与原始，未失真的图像进行对比(也就是, 真值（ground truth))。这个平台将结果以两两配对的方式呈现给研究参与者，让他们在每一对图片中选出一个视觉质量更好的。为了保证参与者做出的是思考后的选择，平台还会让他们在真值图片和图像修复范例结果之间进行选择来验证。 如果应答者没有在一个或两个验证问题中选择出正确答案，平台会将他的所有答案抛弃。最终，平台一共收集到了来自 215 名参与者的 6,945 个成对判断。 |
There was a problem hiding this comment.
Suggested change
| 我们使用[Subjectify.us](http://www.subjectify.us)平台将三个专业艺术家和自动图像修复方法的结果与原始，未失真的图像进行对比(也就是, 真值（ground truth))。这个平台将结果以两两配对的方式呈现给研究参与者，让他们在每一对图片中选出一个视觉质量更好的。为了保证参与者做出的是思考后的选择，平台还会让他们在真值图片和图像修复范例结果之间进行选择来验证。 如果应答者没有在一个或两个验证问题中选择出正确答案，平台会将他的所有答案抛弃。最终，平台一共收集到了来自 215 名参与者的 6,945 个成对判断。 | |
| 我们使用[Subjectify.us](http://www.subjectify.us)平台将三个专业艺术家和自动图像修复方法的结果与原始、未失真的图像（即真值（ground truth））进行对比。这个平台将结果以两两配对的方式呈现给研究参与者，让他们在每一对图片中选出一个视觉质量更好的。为了保证参与者做出的是思考后的选择，平台还会让他们在真值图片和图像修复范例结果之间进行选择来验证。 如果应答者没有在一个或两个验证问题中选择出正确答案，平台会将他的所有答案抛弃。最终，平台一共收集到了来自 215 名参与者的 6,945 个成对判断。 |
|
|
||
|  | ||
|
|
||
| Another surprising result is that the neural method **Generative Image Inpainting**, which was proposed in 2018, scored lower than a non-neural method proposed 14 years ago (**Exemplar-Based Image Inpainting**): | ||
| 另一个令人惊讶的结果是，2018 年提出的神经网络方法 **Generative Image Inpainting**，得分比 14 年前提出的非神经网络方法(**Exemplar-Based Image Inpainting**)还低： |
There was a problem hiding this comment.
Suggested change
| 另一个令人惊讶的结果是，2018 年提出的神经网络方法 **Generative Image Inpainting**，得分比 14 年前提出的非神经网络方法(**Exemplar-Based Image Inpainting**)还低： | |
| 另一个令人惊讶的结果是，2018 年提出的神经网络方法 **Generative Image Inpainting**，得分比 14 年前提出的非神经网络方法（**Exemplar-Based Image Inpainting**）还低： |
|
|
||
|  | ||
|
|
||
| Deep learning has had mind-blowing success in computer vision and image processing over the past few years. For many tasks, deep-learning methods have outperformed their handcrafted competitors in delivering similar or even better results than human experts. For example, GoogleNet’s performance on the ImageNet benchmark exceeds human performance ([Dodge and Karam 2017](https://arxiv.org/abs/1705.02498)). In this post, we compare professional artists and computer algorithms (including recent approaches based on deep neural networks, or DNNs) to determine which can produce better image-inpainting results. | ||
| 过去几年，深度学习在计算机视觉和图像处理领域取得了令人惊艳的成功。许多任务中，深度学习方法不仅能取得与人类专家相近或者更好的结果，还比人工算法有更好的性能表现。比如说 GoogleNet 在 ImageNet 基准上的表现超过了人类([Dodge and Karam 2017](https://arxiv.org/abs/1705.02498))。这篇文章中，我们对专业艺术家和计算机算法（包括最近基于深度神经网络的方法，或者叫 DNN）进行比较，看看哪一方能够在图像修复上取得更好的结果。 |
There was a problem hiding this comment.
『许多任务中，深度学习方法不仅能取得与人类专家相近或者更好的结果，还比人工算法有更好的性能表现。』=>『在许多任务中，深度学习方法优于相应的传统方法，能够取得与人类专家相近甚至是更好的结果。』（注：个人认为in后面是在解释深度学习方法是在什么方面优于传统方法的）
|
|
||
| Image inpainting is the process of reconstructing missing parts of an image so that observers are unable to tell that these regions have undergone restoration. This technique is often used to remove unwanted objects from an image or to restore damaged portions of old photos. The figures below show example image-inpainting results. | ||
| 图像修复是对一幅图像丢失部分的重构过程，使得观察者察觉不到这些区域曾被修复。 这种技术通常用于移除图像中不想要的物体，或者是修复老照片上损坏的部分. 下面的图片展示了图像修复结果的例子。 |
There was a problem hiding this comment.
『这种技术通常用于移除图像中不想要的物体，或者是修复老照片上损坏的部分. 』=>『这种技术通常用于移除图像中多余的元素，或者是修复老照片中损坏的部分。』
|
|
||
| Image inpainting is an ancient art that originally required human artists to do the work by hand. But today, researchers have proposed numerous automatic inpainting methods. In addition to the image, most of these methods also require as input a mask showing the regions that require inpainting. Here, we compare nine automatic inpainting methods with results from professional artists. | ||
| 图像修复是一门古老的艺术，最初需要人类艺术家手工作业。但如今，研究人员提出了许多自动修复方法。除了图像本身，大多数方法还需要展示修复区域的遮罩（mask）作为输入。在这里，我们将对九个自动修复方法和专业艺术家进行比较。 |
There was a problem hiding this comment.
『遮罩』=>『掩码』
『在这里，我们将对九个自动修复方法和专业艺术家进行比较。』=>『接下来，我们将对九个自动修复方法和专业艺术家对图像修复的结果进行比较。』
|
|
||
| To create a set of test images, we cut thirty-three 512×512-pixel patches out of photos from a private collection. We then filled a 180×180-pixel square at the center of each patch with black. The task for both the artists and the automatic methods was to restore a natural look to the distorted image by changing only the pixels in the black square. | ||
| 为了创建测试图片数据集，我们从一个私人照片集中截取了 33 个 512x512 像素的图像片。然后将一个 180x180 像素的黑色正方形填充到每个图像片的中心。而交给艺术家和自动方法的任务是，通过只改变黑色正方形中的像素，来恢复失真图像的自然表现。 |
There was a problem hiding this comment.
『图像片』=>『图像』
『而交给艺术家和自动方法的任务是，通过只改变黑色正方形中的像素，来恢复失真图像的自然表现。』=>『艺术家和自动方法的任务是，通过只改变黑色正方形中的像素，来恢复失真图像的原样。』
|
|
||
| We used a private, unpublished photo collection to ensure that the artists in our comparison had no access to the original images. Although irregular masks are typical in real-world inpainting, we stuck with square masks at the center of the image, since they’re the only type that some DNN methods in our comparison allow. | ||
| 我们使用一个私有，未公开的照片集来保证参与对比的艺术家们没有接触过原始图片。尽管不规则的遮罩是现实世界图像修复的典型特征，但我们不得不在图像中心使用正方形的遮罩，因为它们是我们对比中一些 DNN 方法所唯一允许的类型。 |
|
|
||
| Our study of automatic image-inpainting methods versus professional artists allows us to draw the following conclusions: | ||
| 我们对自动图像修复方法对抗专业艺术家的研究允许我们得到下面的结论： |
There was a problem hiding this comment.
『我们对自动图像修复方法对抗专业艺术家的研究允许我们得到下面的结论：』=>『我们从自动图像修复方法与专业艺术家的对比研究中得到如下结论：』
|
|
||
| We have shared all images and subjective scores collected during the experiment, so you can do your own analysis of this data. | ||
| 我们已经将所有的图片和实验中收集到的主观分数分享出来，因此你可以自己分析这些数据。 |
There was a problem hiding this comment.
『我们已经将所有的图片和实验中收集到的主观分数分享出来，』=>『我们已将实验中收集的图片和主观分数进行了分享』
| 1. 艺术家的图像修复仍然是取得接近真值质量的唯一方法. | ||
| 2. 只有在特定的图像上，自动图像修复方法的结果才能和人类艺术家相媲美。 | ||
| 3. 尽管在这些自动方法中一个深度学习算法取得了第一名，但非神经网络算法仍然处在一个强有力的位置，并且在众多测试的表现超过了深度学习方法。 | ||
| 4. 非深度学习方法和专业艺术家（废话）可以修复任意形状的区域，而大部分基于神经网络的却受到遮罩形状的严格限制。这个约束使得这些方法在现实世界中的适用性变窄了。我们因此突出强调 **Image Inpainting for Irregular Holes Using Partial Convolutions** 这一深度学习方法上，它的开发人员关注于支持任意形状的遮罩。 |
|
|
||
| We have shared all images and subjective scores collected during the experiment, so you can do your own analysis of this data. | ||
| 我们已经将所有的图片和实验中收集到的主观分数分享出来，因此你可以自己分析这些数据。 | ||
|
|
||
| * [Images used in the comparison](https://github.com/merofeev/image_inpainting_humans_vs_ai) |
There was a problem hiding this comment.
『Images used in the comparison』=>『实验对比中的图像』
|
|
||
| We have shared all images and subjective scores collected during the experiment, so you can do your own analysis of this data. | ||
| 我们已经将所有的图片和实验中收集到的主观分数分享出来，因此你可以自己分析这些数据。 | ||
|
|
||
| * [Images used in the comparison](https://github.com/merofeev/image_inpainting_humans_vs_ai) | ||
| * [Subjective scores (including per-image scores)](http://erofeev.pw/image_inpainting_humans_vs_ai/) |
There was a problem hiding this comment.
『Subjective scores (including per-image scores)』=>『主观分数（包含每幅图片的分数）』
|
感谢 @Amberlin1970 @lsvih 辛苦校对，已经做出修改了 |
|
@Glowin 麻烦审核下啦 |
lsvih
approved these changes
Feb 12, 2020
|
我直接在里面改了一点标点符号，您发表文章的时候记得从库里更新一下哈 |
|
@Starry316 已经 merge 啦~ 快快麻溜发布到掘金然后给我发下链接，方便及时添加积分哟。 掘金翻译计划有自己的知乎专栏，你也可以投稿哈，推荐使用一个好用的插件。 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
图像修复：人类和 AI 的对决
第一次翻译~希望校对大大们多给些建议
译文翻译完成，resolve #6682