Learning Degradation Representations for Image Deblurring

In various learning-based image restoration tasks, such as denoising and super-resolution, the degradation representations were widely used to model process handle complicated patterns. However, they are less explored in deblurring blur kernel estimation cannot perform well real-world challenging cases. We argue that it is particularly necessary for since blurry patterns typically show much larger variations than noisy or high-frequency textures. this paper, we propose a framework learn spatially adaptive of images. A novel joint reblurring learning presented improve expressiveness representations. To make learned effective deblurring, Multi-Scale Degradation Injection Network (MSDI-Net) integrate them into neural networks. With integration, MSDI-Net can adaptively. Experiments on GoPro RealBlur datasets demonstrate our proposed with outperforms state-of-the-art methods appealing improvements. The code released at https://github.com/dasongli1/Learning_degradation .