There are different ways to remove or reduce noise, and there are also ways of removing watermarks from images. But Nvidia has recently introduced a deep learning-based approach which has learned to fix photos by looking only at corrupted photos.
The researchers from NVIDIA, Aalto University, and MIT have worked on this project. Similar work in this field has been focused on training neural networks by showing both clean and corrupted images as examples. But in this method, the approach is different. The researchers only used two input images with noise or grain to train the AI to fix them.
While this approach can be useful in photography, I find it interesting that it’s also potentially useful in medicine. The method can be used to enhance noisy MRI images, which can certainly make these images clearer and help doctors see MRI images in more detail.
While noise reduction is definitely a welcomed feature and can be used in the fields other than photography, the same method can also remove text from images. To train the neural network, the researchers used images corrupted with large, varying number of random strings of letters in random places. They were also on top of each other, and with the randomized font size and color.
While the result is pretty impressive, this method can potentially be misused to remove watermarks from images. On the other hand, there are solutions that help to make watermarks difficult to remove. Or perhaps an invisible digital watermarking system can keep your photos safe from theft and watermark removal. At least for now.