Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring

TL;DR

To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture, the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images.

Abstract

Coarse-to-fine schemes are widely used in traditional single-image motion deblur; however, in the context of deep learning, existing multi-scale algorithms not only require the use of complex modules for feature fusion of low-scale RGB images and deep semantics, but also manually generate low-resolution pairs of images that do not have sufficient confidence. In this work, we propose a multi-scale network based on single-input and multiple-outputs(SIMO) for motion deblurring. This simplifies the complexity of algorithms based on a coarse-to-fine scheme. To alleviate restoration defects impacting detail information brought about by using a multi-scale architecture, we combine the characteristics of real-world blurring trajectories with a learnable wavelet transform module to focus on the directional continuity and frequency features of the step-by-step transitions between blurred images to sharp images. In conclusion, we propose a multi-scale network with a learnable discrete wavelet transform (MLWNet), which exhibits state-of-the-art performance on multiple real-world deblurred datasets, in terms of both subjective and objective quality as well as computational efficiency.

Figures (8)

• Figure 1: Performance comparison on the RealBlur-J rim2020real test dataset in terms of PSNR and GMACs. Our proposed MLWNet achieves superiority in comparison with other state-of-the-arts.
• Figure 2: The overall architecture of the proposed MLWNet, the SEB is a simple module designed with reference chen2022simple, the WFB and WHB apply the LWN that implements the learnable 2D-DWT. In training phase, supervised learning is performed using $\mathcal{L}_{multi}$ and self-supervised restraint of the wavelet kernel is performed using $\mathcal{L}_{wavelet}$. In testing phase, only the highest scale restored images is output.
• Figure 3: (a)The process of learnable 2D-wavelet convolution. (b)The construction process of the $N \times N$ 2D-wavelet kernel.
• Figure 4: Visual comparisons on the RealBlur-J dataset rim2020real. The proposed method generates an image with clearer characters.
• Figure 5: Visual comparisons on the RSBlur dataset rim2022realistic. The deblurring performance of the proposed method in low-light is impressive, The recovery of characters and texture structures far exceeds other advanced methods.