Enhance-Then-Deblur: Revealing the Sequential Inductive Bias in Joint Low-Light Image Enhancement and Deblurring
Abstract
Low-light imaging suffers from coupled degradations of low illumination, motion blur, and defocus blur. Recently, cutting-edge solutions suggest employing a single encoder-decoder network for joint image restoration. In this paper, we provide theoretical analysis and experimental evidence to reveal the sequential inductive bias inherent in this joint enhance-and-deblur task. The enhance-then-deblur paradigm consistently outperforms its reverse counterpart across all compared state-of-the-art (SOTA) models. Motivated by this insight, this paper proposes a dual-domain synergistic network (DDSNet) that jointly restores low-light images with a single encoder-decoder network. The proposed model processes global and local features separately in the spatial and frequency domains during the encoding (illumination enhancement) and decoding (motion and defocus deblurring) stages. Critically, a novel frequency-domain alignment bridge (FAB) is introduced to better preserve the frequency features between the encoder and the decoder. Compared with SOTA methods, DDSNet achieves superior performance in terms of PSNR, SSIM, NRQM, and NIQE on the LOL-Blur and Real-LOL-Blur datasets, indicating the efficacy of exploring sequential inductive bias for joint low-light image enhancement and deblurring. Code is publicly available.