Chipmaker NVIDIA has developed an artificial intelligence capable of restoring photos by determining what should be present in white or corrupt spaces. Farewell the duplication stamp of Photoshop!
Global supplier of graphics processors and graphics cards, NVIDIA has already developed an AI with the ability to create fake videos that look like reality, which revealed at the end of 2017. Now the company is creating new buzz with the unveiling of an AI far exceeding the existing possibilities for image editing.
If on classic editing software such as Photoshop and its competitors, it is possible to duplicate adjacent pixels to fill a blank or a corrupted part, it is nonetheless impossible to determine what should be in the right place. This obviously concern old photos whose colors can disappear and make disappeared details impossible to restore.
The AI developed by NVIDIA , published on arXiv.org, is a deep learning system that can determine what should be visible in corrupted parts and white spaces. The method used here is partial convolution and an AI trained to reconstruct photos among tens of thousands on which variation holes were generated.
“We are the first to demonstrate the effectiveness of deep-painting models of image in irregularly shaped holes,” said the researchers at the American firm.
Moreover, if the results are not always convincing, as shown in the press release published by NVIDIA vidia on April 22, 2018, artificial intelligence could be improved and become a reference in image editing. Finally, NVIDIA will present its concept at the International Conference on Learning Representations (ICLR) which is currently taking place, from April 30 to May 3, 2018 at the Vancouver Convention Center (Canada).