Multi–Grid Back–Projection Networks

Multi–grid back–projection (MGBP) is a fully–convolutional network architecture that can learn to restore images and videos with upscaling artifacts. Using the same strategy of multi–grid partial differential equation (PDE) solvers this multiscale scales computational complexity efficiently increasing output resolutions. The basic processing block inspired in iterative (IBP) algorithm constitutes type cross–scale residual feedback from low resolution references. performs par state–of–the-arts alternatives for regression targets aim recover an exact copy high image or video which only downscale known. A perceptual quality target aims create more realistic outputs by introducing artificial changes be different original content as long they are consistent input. For we propose using noise inputs control amount details generated output. input controls innovation uses details. effectiveness shown benchmarks it explained particular traverse perception–distortion plane.