Blind Super-Resolution with Kernel-Aware Feature Refinement

Abstract

Numerous learning-based super-resolution (SR) methods that assume the blur kernel is known in advance or hand-crafted perform excellently on the synthesized images. While these methods fail to produce satisfactory results when applied to real-world images. Therefore, several blind SR methods are proposed, where the blur kernel is unavailable. However, they only consider the Gaussian blur kernel, which is not enough for real-world images. To address this problem, we propose to synthesize training data with real blur kernels estimated from real sensor images. Further, we find that models trained with the data synthesized with the estimated real blur kernels still cannot perform well enough without awareness of the blur kernel. Thus, we first propose a new kernel predictor (KP) to predict the blur kernel from the input low-resolution image. Then we design the kernel-aware feature refinement (KAFR) module to utilize the predicted blur kernel to facilitate the restoration of the high-resolution image. Extensive experiments on synthesized and real-world images show that the proposed method achieves superior performance against the state-of-the-art (SOTA) methods both on effectiveness and efficiency.