Computer Vision News - January 2022

4 Computer Vision Research Resolution-robust Large Mask Inpainting with Fourier Convolutions Welcome to our last issue before the year, which for a lot of people comes to a change During our articles this year we looked at topics ranging from neuroscience (I hope that you enjoyed especially the last article, which was a great introductory but also more in-depth programmatic approach to neuroimaging), to liver, cardiac disease and MR angiography, deep learning techniques as applied to biological imaging and medical imaging, to many many more. What was your favorite? What does excite you? Feel free to contact and let us know of new and interesting topics that you enjoy, or you would like to see covered in the usual channels (social above or directly on Computer Vision News)! The pandemic seems to have a come-back but this time with hopefully more vaccinated people and less losses. Let’s hope that it will be normality, as with the flu and we’ll learn to be more thoughtful and careful. For everyone celebrating, have an amazing new year! Review This month, we’ll present “Resolution-robust Large Mask Inpainting with Fourier Convolutions” from Roman Suvorov and colleagues in Samsung Research, Skolkovo Institute of Science and Technology, Moscow, Russia and EPFL in Lausanne. We thank the authors (special thanks to Arsenii Ashukha) for authorizing Computer Vision News to use photos and extra material! I personally need to thank Ralph, for his amazing support, suggestion and help contacting me with his vast network! The targetof thisarticleandtherecommendedapproach is toattempt andsolve the image inpainting problem. I can hear you asking already, whattt? This is simply the act of filling the missing parts of an image with a “realistic” approach. It is important to “understand” the large-scale structure of natural images and to perform image synthesis. In the pre-deep learning times, this has been extensively studied (you can explore more about this in the references of the original paper). Now how is this going to be achieved? The plan is to utilize Fourier convolutions , using a single- stage deep learning network. The method which was used by the authors to achieve this goal is to inpaint a color image x masked by a binary mask of unknown pixels. The mask m is stacked with the masked image x’ m, resulting in a four-channel input tensor. A feed-forward inpainting network was used, named f θ (•), which is further referred around as the “generator” .