MHRNet: A Multi-stage Image Deblurring Approach with High-Resolution Representation Learning

Image deblurring is a classic task in computer vision that aims to recover clean images from blurred images and is widely used in surveillance, medical imaging and so on. Existing methods typically apply a multi-stage encoder-decoder architecture to learn features at different scales, and they have achieved remarkable results. However, these methods usually map the input to a low-resolution (LR) image to expand its receptive field and then gradually reverse this image to the original resolution. Although these approaches obtain rich semantic information via spatial resolution reduction, they lose a large amount of spatial information, which is essential for image deblurring and extremely difficult to recover. To solve this problem, we propose a novel multi-stage model with high-resolution (HR) representation learning (MHRNet). In this model, HR representations are always preserved to reduce the loss of spatial information, and the features across all the scales at each resolution are fused to obtain spatially accurate and semantically rich features. Extensive experiments conducted on the GoPro, HIDE and RealBlur datasets demonstrate that MHRNet outperforms the state-of-the-art (SOTA) methods and reaches a 33.99-dB peak signal-to-noise ratio (PSNR) on the GoPro dataset.