检测图像是否有水印？
我们的网站允许人们上传图像。然而，我们不允许带水印的图像，但用户仍然上传许多水印图像。是否有一些软件/代码可以（至少在大多数情况下）捕获确实带有水印的图像，例如徽标/图像？我不确定是否有某种标准。
Our website allows people to upload images. However, we don't allow watermarked images, yet many do still get uploaded by users. Is there some software/code that can (at least in most cases) catch images that do have watermarks such as logos/images? I'm not sure if there is some sort of a standard.
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（4）
您可以通过图像分类来做到这一点。
基本上，通过输入一些带有水印的图像和一些不带水印的图像来训练CNN（卷积神经网络）模型，然后使用该模型来判断任何新图像中水印的概率。
您可以将迁移学习应用于一些现有的预训练模型（截至目前，inception v3 是最好的），这些模型可以针对您的特定分类目的进行重新训练。
例如，此链接展示了如何识别图像是向日葵、雏菊还是玫瑰的图像。
https://www.tensorflow.org/tutorials/image_retraining
这是一个 5 分钟的快速教程，内容涉及构建张量流图像分类器：https://youtu.be/QfNvhPx5Px8
You can do it via image classification.
Basically, train a CNN(Convolutional neural Network) model by feeding in some images with watermark and some without watermark in it and then use this model to judge the probability of watermark in any new image.
You can apply transfer learning on some existing pre-trained models(as of today inception v3 is the best out there) which can be retrained for your specific classification purpose.
For example this link shows how to do it to identify whether an image is that of a sunflower or a daisy or a rose.
https://www.tensorflow.org/tutorials/image_retraining
Here is a quick 5 minute tutorial about building a tensorflow image classifier: https://youtu.be/QfNvhPx5Px8
要检测图像上的任何类型的徽标都会非常复杂。你需要类似于面部识别的东西，以及大量的人工智能……
为了使其相当高效，你需要一个徽标库来查找并知道它们应用于图像的何处。如果徽标始终位于同一位置，您只需遮盖其所在位置的像素，然后计算它与徽标像素的接近程度。如果徽标的大小和位置不同，事情就会变得更加复杂。
To detect any kind of logo on an image would be quite complicated. You would need something similar to face recognition, and a lot of AI...
To make it reasonably efficient you would need a library of logos to look for, and know where they are applied on the images. If the logo is always in the same place, you could just mask out the pixels where it would be, and calculate how close it is to the pixels of the logo. If logos varies in size and position, it gets more complicated.
您无法自动检测水印。最好的办法是让其他人能够轻松报告带有水印的图像，并且一旦报告，将它们置于保留状态，直到验证它们是否带有水印为止，它们不会显示。
You can't automatically detect a watermark. The best thing to do is make it real easy for others to report images that have a watermark and once reported, put them in a holding state where they aren't displayed until it's verified they either do or don't have a watermark.
对于某种人工智能来说，这是可能的，至少有一定的可能性。
更准确地说，只要您可以定义水印是什么，这是可能的，
这是最大的问题。通用水印检测实际上无法检测到，
考虑照片等广告牌上的徽标。
With certain kind of AI it would be possible, at least with certain probability.
More precisely said it IS possible provided that you CAN define what the watermark is,
which is the greatest problem. Generic watermark detection is virtually undetectable,
consider logo at billboard at photo etc.