Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Making Visible Watermarks More Effective

research.googleblog.com

Post reply on HN
Print view View original

21–30 of 98 posts

Re: Making Visible Watermarks More Effective

#21

Post by Ajedi32 »

dtech wrote: post #15
Ajedi32 wrote: post #13

I wonder if removing watermarks from images would be something a machine learning algorithm would be well suited for. Admittedly my understanding of machine learning is rather limited, but it seems like it'd be pretty easy to generate a large training set of watermarked and unwatermarked images for the algorithm to train on. No idea how effective it'd be though.

If you read the blog post, they link to their own research that does exactly that.

Could you show me where? I saw nothing in that article mentioning machine learning or deep learning specifically, and the paper the blog post is based on seems to use non-ML based techniques.

My assumption is that a deep learning algorithm trained on multiple different styles and variations of watermarked images would be much more robust to the sort of changes Google proposed in this blog post as a way to defeat their existing algorithms.

Re: Making Visible Watermarks More Effective

#24

Post by sp332 »

bwang29 wrote: post #11

I think the question is why would we need protective watermark anyway if stock photo companies are already crawling and sometimes phishing for use of licensed stock photography on the web and then directly send out an DCMA or a charge? I've been in many situations where the copyright owners reached out for damage fees after downloading a full-res, un-watermarked photo from free stock photos sites in blog posts, so I'…

This is extremely time-consuming. It's enough to put smaller operations out of business. https://arstechnica.com/tech-policy/2014/09/one-mans-endless...

Re: Making Visible Watermarks More Effective

#25

Post by sp332 »

Ajedi32 wrote: post #13

I wonder if removing watermarks from images would be something a machine learning algorithm would be well suited for. Admittedly my understanding of machine learning is rather limited, but it seems like it'd be pretty easy to generate a large training set of watermarked and unwatermarked images for the algorithm to train on. No idea how effective it'd be though.

Not all machine learning is "deep neural networks". I think that's why your comment is being downvoted. The video shows specific model parameters that are learned for each watermark.

Re: Making Visible Watermarks More Effective

#26

Post by icebraining »

nxsynonym wrote: post #4

Thinking out loud here - but would this be a good use case for blockchain technology? The problem with visible watermarks is it detracts from the image visibility. Nobody wants to look at photos or digital art pieces with huge ugly watermarks on them. Could blockchain tech help establish ownership in a way that would make watermarking obsolete? edit: cool - downvotes for asking a question. Real nice guys.

cool - downvotes for asking a question. Real nice guys.

I didn't downvote you, but I suspect at this point people are a bit fed up of the constant suggestions of using the Blockchain for everything.

Re: Making Visible Watermarks More Effective

#27

Post by jotato »

lightbyte wrote: post #20
jotato wrote: post #12

Earlier quoted context omitted.

I wonder what would happen if I took 1000 copies of the same image that is using a warped mark and ran it through the algorithm....

The algorithm is calculating the mean of the set of images, so putting 1000 copies of the same image will do absolutely nothing

i mean to remove the watermark on that specific image

Re: Making Visible Watermarks More Effective

#28

Post by Retric »

pps43 wrote: post #9
kuschku wrote: post #2

This is just adding a warping effect, and I’m quite sure that if this technology had already existed, then the same team at Google that did this research would have, with a similar amount of work, been able to circumvent this technique, too. I mean, de-warping warped imagery is something that Google’s image stabilization software used on YouTube can already do very well. Adapting it for this purpose should be possibl…

If the same watermark is added to each image, then you can estimate the watermark by averaging many images and subtract it to get original image back. If each image has a slightly different watermark, then simple averaging won't work. Instead you need to come up with a model that describes how the watermark is changing, then estimate parameters of that model. The more complex the change, the more images you need for…

Averaging will still provide the average watermark given a large enough sample set. Which is then a single geometric transformation from the used watermark. Estimating that transformation for a specific image should be fairly straightforward as they could easily reverse a watermark that was offset. At which point it's reversible.

Re: Making Visible Watermarks More Effective

#30

Post by sly010 »

throwasehasdwi wrote: post #5
nxsynonym wrote: post #4

Thinking out loud here - but would this be a good use case for blockchain technology? The problem with visible watermarks is it detracts from the image visibility. Nobody wants to look at photos or digital art pieces with huge ugly watermarks on them. Could blockchain tech help establish ownership in a way that would make watermarking obsolete? edit: cool - downvotes for asking a question. Real nice guys.

Probably not but I was thinking either ML or VR could help

You meant AI...
Post reply on HN