首页|期刊导航|四川轻化工大学学报(自然科学版)|低光图像增强方法研究现状及展望

低光图像增强方法研究现状及展望OA

Research Status and Future Prospects of Low-Light Image Enhancement Methods

中文摘要英文摘要

低光图像受到光照不足、传感器噪声、颜色偏移和ISP非线性影响,常出现亮度压缩、暗区细节淹没、噪声放大和结构失真等退化问题,并削弱检测、识别、分割等下游任务.针对既有综述多沿用"传统方法-Retinex模型-深度学习方法-生成模型"分类而问题机制分析不足的问题,本文提出"退化主因-建模约束-应用目标"三维框架,将低光增强方法归纳为曝光响应重映射型、照明-反射可解释分解型、噪声-复合退化感知型、结构-频率协同重建型、生成先验与语义可控型、应用约束型6类,并讨论真实/合成数据域偏移、RAW/sRGB成像域适配差异,以及客观指标、主观观感与下游任务性能之间的评价不匹配问题.最后,从基础模型、多模态协同、端侧部署、隐私保护和任务驱动评价等方面展望图像增强发展趋势,为低光图像增强研究提供问题导向、应用导向和可解释性更强的综述框架.

Low-light images are affected by insufficient illumination,sensor noise,color deviation,and nonlinear ISP processing,which commonly lead to brightness compression,loss of dark-region details,noise amplification and structural distortion,and further weaken downstream tasks such as detection,recognition and segmentation.Existing surveys often follow the taxonomy of traditional methods,Retinex-based models,deep learning methods and generative models,which is insufficient to reveal the relationships among degradation mechanisms,modeling constraints,and application requirements.This paper proposes a degradation-constraint-application(DCA)framework and reorganizes existing low-light image enhancement methods into six categories:exposure-response remapping,interpretable illumination-reflectance decomposition,noise-coupled degradation-aware restoration,structure-frequency collaborative reconstruction,generative-prior and semantically controllable enhancement,and application-constrained enhancement.The paper further discusses domain gaps between real and synthetic data,adaptation differences between RAW and sRGB domains,and the inconsistency among objective metrics,subjective perception,and downstream task performance.Finally,recent trends are analyzed from foundation models,multimodal modeling,edge deployment,privacy-preserving enhancement and task-driven evaluation.The proposed framework provides a problem-oriented,application-aware,and interpretable perspective for understanding the research status and future prospects of low-light image enhancement.

谢乐惠;周顺勇;刘盈盈;车召旭;林伟;梁蔚

四川轻化工大学 自动化与信息工程学院,四川 宜宾 644000||智能感知与控制四川省重点实验室,四川 宜宾 644000四川轻化工大学 自动化与信息工程学院,四川 宜宾 644000||智能感知与控制四川省重点实验室,四川 宜宾 644000四川轻化工大学 自动化与信息工程学院,四川 宜宾 644000||智能感知与控制四川省重点实验室,四川 宜宾 644000四川轻化工大学 自动化与信息工程学院,四川 宜宾 644000||智能感知与控制四川省重点实验室,四川 宜宾 644000四川轻化工大学 自动化与信息工程学院,四川 宜宾 644000||智能感知与控制四川省重点实验室,四川 宜宾 644000四川轻化工大学 自动化与信息工程学院,四川 宜宾 644000||智能感知与控制四川省重点实验室,四川 宜宾 644000

信息技术与安全科学

低光图像增强DCA框架退化建模Retinex视觉Transformer生成模型任务驱动增强真实世界泛化

low-light image enhancementDCA frameworkdegradation modelingRetinexvision Transformergenerative modelstask-driven enhancementreal-world generalization

《四川轻化工大学学报(自然科学版)》 2026 (4)

58-73,16

国家自然科学基金项目(61801319)四川省科技厅项目(2020YFSY027)

10.11863/j.suse.2026.04.06

评论