GAFL: Global adaptive filtering layer for computer vision

We devise a universal global adaptive filtering layer, GAFL, capable of “learning” optimal frequency filter for each image in dataset together with the weights base neural network that performs some computer vision task. The proposed approach takes source spatial domain, selects best frequencies Fourier domain benefit task, and prepends inverse-transform to main joint training. Remarkably, such simple add-on optimizing content an input specific dramatically improves performance regardless its design. observe light networks gain noticeable boost metrics; whereas, training heavy ones converges faster when GAFL is prepended architecture. showcase layer four classical tasks: classification, segmentation, denoising, erasing, considering popular natural medical data benchmarks.