2021/08/11 by Jingyun Liang, Liang, Jingyun, Guolei Sun +7 · 3 citations
Computer Science · Engineering · #Advanced Image Processing Techniques #Image Processing Techniques and Applications #Advanced Vision and Imaging
paper · pdf · doi:10.48550/arxiv.2108.05302
Existing blind image super-resolution (SR) methods mostly assume blur kernels\nare spatially invariant across the whole image. However, such an assumption is\nrarely applicable for real images whose blur kernels are usually spatially\nvariant due to factors such as object motion and out-of-focus. Hence, existing\nblind SR methods would inevitably give rise to poor performance in real\napplications. To address this issue, this paper proposes a mutual affine\nnetwork (MANet) for spatially variant kernel estimation. Specifically, MANet\nhas two distinctive features. First, it has a moderate receptive field so as to\nkeep the locality of degradation. Second, it involves a new mutual affine\nconvolution (MAConv) layer that enhances feature expressiveness without\nincreasing receptive field, model size and computation burden. This is made\npossible through exploiting channel interdependence, which applies each channel\nsplit with an affine transformation module whose input are the rest channel\nsplits. Extensive experiments on synthetic and real images show that the\nproposed MANet not only performs favorably for both spatially variant and\ninvariant kernel estimation, but also leads to state-of-the-art blind SR\nperformance when combined with non-blind SR methods.\n