GAFL: Global adaptive filtering layer for computer vision

摘要

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