Answers are grounded in this paper and cite their sources.
Low-light image enhancement (LLIE) can effectively restore the quality of photos captured in suboptimal illumination conditions. But most current LLIE methods fail to balance color correction and structure-texture boosting. To tackle these challenges, we propose a Transformer-based approach constrained by the color and structure-texture priors called CSPFormer. Specifically, we introduce a color prior generator that leverages color distribution features as supplementary in-formation for global attention, and a color bank strategy to capture color dynamics. Concurrently, we introduce a structure-texture prior generator that applies localized attention guided by window-level variance selection to disentangle and extract distinct structure and texture priors. By employing these specialized priors with different Transformer attentions, we extend the capabilities of Transformer methods beyond conventional long- and short-range feature capture. Our enhanced results are characterized by continuous vivid color and clearer details. Extensive experiments on various datasets demonstrate that our approach achieves competitive performance compared with representative recent LLIE methods. The code is available at https://github.com/harryxin1511/CSPFormer.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: