Customized Treatment Per Pixel for Blind Image Super-Resolution

Abstract

Blind image super-resolution task aims at restoring high-resolution images from their low-resolution counterparts by reversing the unknown degradation. Existing methods have achieved promising results when handling degradation with isotropic or anisotropic gaussian blur, whereas suffer from performance drop when addressing degradation with motion blur. Compared with gaussian blur, motion blur is more diverse, i.e., each pixel moves a peculiar distance in distinct orientation, thereby each pixel requires individual treatment. To tackle this degradation with motion blur issue, we propose a novel blind image super-resolution method named deformAble receptive Super Resolution (ArcSR), which provides deformable receptive field and unique parameters for each pixel. Specifically, we propose Deformable Mutual convolution (DMconv) and Kernel Guided convolution (KGconv) for blur kernel estimation and super-resolution, respectively. The DMconv explores the correlation within channels of image features and achieves deformable receptive field by redesigning deformable convolution to generate kernels for each pixel. Meanwhile, the KGconv views the estimated kernel as an attention matrix for convolutional parameters and gives each pixel disparate convolutional parameters to rewrite the missing high-frequency information. Comprehensive experiments demonstrate the superiority of our method.