LGScanNet: Gated Local–Global Selective Scanning for Joint Low-Light Enhancement and Deblurring
TL;DR
LGScanNet, a gated local–global selective scanning network for joint low-light enhancement and deblurring, is proposed and controlled comparisons against parameter-matched convolutional and state-space controls indicate that the performance gains cannot be explained by increased model capacity or by selective scanning alone.
Abstract
Images captured in low-light environments often suffer from coupled degradations, including insufficient illumination, amplified noise, and motion blur caused by long exposure: low illumination weakens the structural cues needed for deblurring, while blur further disperses already degraded edges and textures. Existing decoders for this joint restoration problem remain largely convolutional, propagating distant blur information only indirectly through repeated local operations, whereas directly substituting a state-space or attention module risks overwriting locally reliable structure with global context. This motivates a decoder that preserves local detail reconstruction while selectively admitting long-range directional dependencies. We propose LGScanNet, a gated local–global selective scanning network for joint low-light enhancement and deblurring. LGScanNet retains the illumination-oriented encoder of DarkIR and redesigns its deblurring decoder around the proposed Local–Global Deblurring Block (LGDB), which combines a Di-SpAM-based local detail branch with an Affine-Calibrated Multi-Head Selective Scan (AC-MHSS) branch. In AC-MHSS, shared affine calibration conditions four independently parameterized directional scan heads before they model direction-dependent blur trajectories. The local and global representations are then combined through Local-Guided Global Injection (LGGI), which anchors the fused feature on the local branch and admits the global response only through a learned gate and a channel-wise scale initialized near zero, so that global context is introduced gradually as a controlled correction to the local representation. Controlled comparisons against parameter-matched convolutional and state-space controls, together with component-wise ablations, indicate that the performance gains cannot be explained by increased model capacity or by selective scanning alone. On LOLBlur-Synthetic, LGScanNet-M improves our reproduced DarkIR-M baseline by approximately 0.70 dB in PSNR.