图像修复：人类和 AI 的对决

TODO1/image-inpainting-humans-vs-ai.md

@@ -2,38 +2,38 @@
 > * 原文作者：[Mikhail Erofeev](https://medium.com/@mikhail_26901)
 > * 译文出自：[掘金翻译计划](https://github.com/xitu/gold-miner)
 > * 本文永久链接：[https://github.com/xitu/gold-miner/blob/master/TODO1/image-inpainting-humans-vs-ai.md](https://github.com/xitu/gold-miner/blob/master/TODO1/image-inpainting-humans-vs-ai.md)
-> * 译者：
+> * 译者：[Starry](https://github.com/Starry316)
 > * 校对者：
 
-# Image Inpainting: Humans vs. AI
+# 图像修复：人类和 AI 的对决
 
 ![](https://cdn-images-1.medium.com/max/6000/1*HQxitL28dDEKe1dPp9wdmQ.png)
 
-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。
 
-## What Is Image Inpainting?
+## 图像修复是什么？
 
-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。
 
-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。
 
-![Image-inpainting example: removing an object. (Image from [Bertalmío et al., 2000](https://conservancy.umn.edu/bitstream/handle/11299/3365/1/1655.pdf).)](https://cdn-images-1.medium.com/max/2152/1*EOuFiCNYdNde05bi9UmB8A.jpeg)
+![图像修复例子： 移除一个物体。 (图片来自 [Bertalmío et al., 2000](https://conservancy.umn.edu/bitstream/handle/11299/3365/1/1655.pdf).)](https://cdn-images-1.medium.com/max/2152/1*EOuFiCNYdNde05bi9UmB8A.jpeg)
 
-![Image-inpainting example: restoring an old, damaged picture. (Image from [Bertalmío et al., 2000.](https://conservancy.umn.edu/bitstream/handle/11299/3365/1/1655.pdf))](https://cdn-images-1.medium.com/max/2412/1*_Ldd9jY-9xS2OEE6Z8FTfw.jpeg)
+![图像修复例子： 修复一张老旧，损坏的照片。 (图片来自 [Bertalmío et al., 2000.](https://conservancy.umn.edu/bitstream/handle/11299/3365/1/1655.pdf))](https://cdn-images-1.medium.com/max/2412/1*_Ldd9jY-9xS2OEE6Z8FTfw.jpeg)
 
-## Data Set
+## 数据集
 
-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。
 
-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。
 
-Below are thumbnails of the images in our data set.
+下面是我们数据集中图片的略缩图。
 
-![Image-inpainting test set.](https://cdn-images-1.medium.com/max/3188/1*_sOFyA9XY3ATpW4aGdnTtA.png)
+![图像修复测试集。](https://cdn-images-1.medium.com/max/3188/1*_sOFyA9XY3ATpW4aGdnTtA.png)
 
-## Automatic Inpainting Methods
+## 自动修复方法
 
-We applied to our test data set six inpainting methods based on neural networks:
+我们对测试数据集使用六种基于神经网络的图像修复方法：
 
 1. Deep Image Prior ([Ulyanov, Vedaldi, and Lempitsky, 2017](https://arxiv.org/abs/1711.10925))
 2. Globally and Locally Consistent Image Completion ([Iizuka, Simo-Serra, and Ishikawa, 2017](http://hi.cs.waseda.ac.jp/~iizuka/projects/completion/en/))
@@ -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：
 
 1. Exemplar-Based Image Inpainting ([Criminisi, Pérez, and Toyama, 2004](http://www.irisa.fr/vista/Papers/2004_ip_criminisi.pdf))
 2. Statistics of Patch Offsets for Image Completion ([He and Sun, 2012](http://kaiminghe.com/eccv12/index.html))
 3. Content-Aware Fill in Adobe Photoshop CS5
 
-## Professional Artists
+## 专业艺术家
 
-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。
 
-Below are results:
+下面是结果：
 
 ![](https://cdn-images-1.medium.com/max/2000/1*tDhUKacPIfjkfdC24tXd_Q.png)
 
-## Humans vs. Algorithms
+## 人类 vs. 算法
 
-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。
 
-Below are the overall and per-image subjective quality scores for this comparison:
+下面是这次比较的总体和每幅图像的主观质量分数：
 
 ![Subjective-comparison results for images inpainted by professional artists and by automatic methods.](https://cdn-images-1.medium.com/max/2852/1*vQFC5lH3mGjAMJyTosgSjw.png)
 
-As the **“Overall”** plot illustrates, the artists all outperform the automatic methods by a large margin. In only one case did an algorithm beat an artist: the **“Urban Flowers”** image inpainted by the non-neural method **Statistics of Patch Offsets** (He and Sun, 2012) received a higher ranking than the image drawn by **Artist #1**. Moreover, only images inpainted by the artists are on a par with or look even better than the original undistorted images: the **“Splashing Sea”** images inpainted by **Artist #2** and **#3** garnered a higher quality score than ground truth, and the **“Urban Flowers”** image inpainted by **Artist #3** garnered a score just slightly lower than ground truth.
+ **“Overall”** 图表表明，所有艺术家的表现都比自动方法好上一大截。只有在一个例子下，一个算法击败了一名艺术家： **Statistics of Patch Offsets** (He and Sun, 2012) 对 **“Urban Flowers”** 图片的修复，得分高过了 **Artist #1** 绘制的图片。还有，只有艺术家修复的图片能够媲美甚至比原图更好：**Artist #2** 和 **#3** 绘制的 **“Splashing Sea”** 图片得到了比真值更高的质量分数，**Artist #3** 绘制的 **“Urban Flowers”** 得分只比真值低一点点。
 
-First place among the automatic approaches went to the deep-learning method Generative Image Inpainting. But it wasn’t a landslide victory, since this algorithm never achieved the best score for any image in our study. First place for **“Urban Flowers”** and **“Splashing Sea”** went to the non-neural methods **Statistics of Patch Offsets** and **Exemplar-Based Inpainting**, respectively, and first place for **“Forest Trail”** went to the deep-learning method **Partial Convolutions**. Notably, according to the overall leaderboard, the other deep-learning methods were outperformed by non-neural ones.
+在自动方法中获得第一名的是深度学习方法 Generative Image Inpainting。但这并不是压倒性的胜利，因为在我们的研究中，这个算法从来没有在任何图片中取得最高分数。对于 **“Urban Flowers”** 和 **“Splashing Sea”** 第一名分别是非神经网络方法的 **Statistics of Patch Offsets** 和 **Exemplar-Based Inpainting**，而 **“Forest Trail”** 的第一名是深度学习方法 **Partial Convolutions**。值得注意的是，根据总体的排行榜来看，其它的深度学习方法都被非神经网络方法超越。
 
-## Interesting Examples
+## 一些有趣的例子
 
-Several results caught our attention. The non-neural method **Statistics of Patch Offsets** (He and Sun, 2012) produced an image that the comparison participants generally preferred over an artist-drawn image:
+一些结果引起了我们的注意。非神经网络的方法 **Statistics of Patch Offsets** (He and Sun, 2012) 产生的图片比起艺术家绘制的图片更受到参与比较者的青睐：
 
 ![](https://cdn-images-1.medium.com/max/2000/1*3dDa-RRW6QhZwiFVIrlnfg.png)
 
-In addition, the image from the top-ranked neural method, **Generative Image Inpainting**, earned a lower score than the non-neural method **Statistics of Patch Offsets**:
+此外，高排名的神经网络方法 **Generative Image Inpainting** 得到的图像，获得了比非神经网络方法 **Statistics of Patch Offsets** 更低的分数：
 
 ![](https://cdn-images-1.medium.com/max/2000/1*aVpvEogJotWTi2F1YjfJvg.png)
 
-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：
 
 ![](https://cdn-images-1.medium.com/max/2000/1*UFvv4H_C1j-F3pVSi5aPlw.png)
 
-## Algorithms vs. Algorithms
+## 算法之间的比较
 
-To further compare neural image-inpainting methods with non-neural ones, we performed an additional subjective comparison using Subjectify.us. Unlike the first comparison, we compared these methods using the entire 33-image data set.
+为了深入比较神经网络方法和非神经网络方法，我们使用 Subjectify.us 进行了一次额外的主观比较。与第一次比较不同，我们使用完整的 33 张图片数据集对这些方法进行比较。
 
-Below are the overall subjective scores computed using 3,969 pairwise judgments from 147 study participants.
+下面是从 147 名研究参与者给出的 3,969 个成对判断中得到总体主观分数。
 
-![Subjective comparison of automatic image-inpainting methods.](https://cdn-images-1.medium.com/max/2358/1*sfhG6AFZ546S6z51aEmuhg.png)
+![自动图像修复方法的主观比较。](https://cdn-images-1.medium.com/max/2358/1*sfhG6AFZ546S6z51aEmuhg.png)
 
-These results confirm our observations from our other comparison. First place (after ground truth) went to **Generative Image Inpainting** trained on the Places2 data set. The **Content-Aware Fill Tool in Photoshop CS5**, which doesn’t use neural networks, was just slightly behind first place. **Generative Image Inpainting** trained on ImageNet took third place. Notably, all other deep-learning approaches were outperformed by non-neural ones.
+这些结果证实了我们从其他比较中得到的观测结果。第一名（在真值之后）是在 the Places2 数据集上训练的 **Generative Image Inpainting**。没有使用神经网络的 **Content-Aware Fill Tool in Photoshop CS5** 以很小的差别位居第一名之后。 在 ImageNet 上训练的 **Generative Image Inpainting** 获得第三名。值得注意的是，其他所有的非神经网络方法的表现都超过了深度学习方法。
 
-## Conclusion
+## 结论
 
-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：
 
-1. Inpainting by artists remains the only way to achieve quality similar to that of ground truth.
-2. The results of automatic inpainting methods can be on a par with those of human artists only for certain images.
-3. Although first place among the automatic methods went to a deep-learning algorithm, non-neural algorithms maintain a strong position and outperformed the deep-learning alternatives on numerous tests.
-4. While non-deep-learning methods and professional artists (of course) can inpaint regions of arbitrary shape, most neural-based methods impose strict limitations on the mask shape. This constraint further narrows the real-world applicability of these methods. We therefore highlight the deep-learning method **Image Inpainting for Irregular Holes Using Partial Convolutions**, whose developers focused on supporting arbitrary masks.
+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 believe future research in this field as well as the growth in GPU computational power and RAM size will allow deep-learning algorithms to outperform their conventional competitors and deliver image-inpainting results on a par with those of human artists. Nevertheless, we emphasize that given the state of the art, choosing a classic image- or video-processing method may be better than blindly choosing a new deep-learning approach just because it’s the latest thing.
+我们相信未来这一领域的研究，以及 GPU 算力和 RAM 容量的增长，将会使得深度学习算法胜过它们的传统竞争者，得到和人类艺术家相媲美的图像修复结果。然而我们强调，在目前最新的技术下，选择一个传统的图像或视频处理方法，也许会比盲目地只是因为新，而选择一个新的深度学习方法更好。
 
-## Bonus
+## 福利
 
-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/)