![]() But I never saw anyone applying this for removing the watermark. Yes, it's not a holy new solution, I've seen it done before in form of image inpainting. If you look at the bigger picture of watermark removal, then, in a nutshell, its just an image inpainting task right? So, all we need to do is, roughly highlight the watermarked region from any paint software and you're good to go. I hope you can see where I'm going from this □. Read the last bold statement again, if we solve that issue, then it's just a matter of following the first scenario ain't it. What exactly is not the part of watermark?Īnd we want to do this without any training!!! Why? Well there's no point, I mean we know for a fact that the generator is indeed capable of inpainting the watermark, it's just us who are not able to provide the answers to the generator for questions above.We can provide absolutely no info to the generator regarding: # Let us see how seriously difficult this is: Even the authors provided the outputs for the first scenario only. For very trivial causes, the first scenario was too easy to tackle than this one. And this is the actual and highly realistic scenario because of obvious reasons. In this scenario, we'll have the watermarked image only. Run $ python inference.py with following arguments. The generator simply takes the random noise with same height and width as of watermarked image, which we can regularize, and produces the outputs. And the authors propose to simply use L2 loss to minimize the distance between them. So if we know what Watermark is, then its just a matter of training a generator that produces outputs, such that Watermarked Image is equal to Generated Image * Watermark. The scale, position, rotation and other spatial transformations of the watermark, exactly matches the applied watermark of the image.Īny watermarked image can be represented as the Hadamard product of Original Image and Watermark. The watermark that is applied to the watermarked image, is available to you. # So in this scenario, the requirements are: In this repo, I've implemented the watermark removal task, and the results are just as good as claimed by the authors. ![]() Thus most of the image restoration tasks, for example, denoising, super-resolution, artefacts removal, watermark removal etc can be done with highly realistic results without any training. ![]() This paper shows that the structure of a generator alone is sufficient to provide enough low-level image statistics without any learning. And it is believed that their great performance is because of their ability to learn realistic image priors from training on large datasets.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |