Semi-Blind Super-Resolution with Kernel-Guided Feature Modification

Non-blind single image super-resolution(SISR) methods have demonstrated their high efficiency and effectiveness on synthesized images. However, these non-blind methods assume that the blur kernel is predefined or hand-crafted, which degrades the performances on real-world images. To deal with real-world images, we propose a semi-blind super-resolution (SR) method where the blur kernel is unknown. We first design a Kernel Predictor(KP) to predict the blur kernel from the input images with a well-designed kernel set. Then we follow the meta-learning and design the Meta Feature Modification(MFM) module to modify the image features. The weight of the modification layer, which is the core component of the MFM module, is generated with the predicted blur kernel. With the generated weight, the modification layer modifies the image features with the guidance of the predicted kernel. Extensive experiments on synthesized and real-world images show that the proposed method achieves superior performance in the blind SR problem.