摘要
暗光场景下拍摄的图像容易受暗光、噪声、模糊等多种退化因素的影响,导致其内容可视度低且视觉观感较差.多退化的暗光图像对现有图像增强方法提出了挑战:一方面,暗光图像增强或去模糊等方法不能涵盖所有的退化类型,而组合使用已有方法的效果受到计算开销增加与误差累积的限制;另一方面,已有多退化暗光图像增强方法采用了先提升亮度再去除模糊的策略,这种顺序处理方式会增加特征线索丢失的风险,不利于细节复原.为应对上述挑战,本文提出渐进式边缘感知交互增强网络(Progressive Edge-aware Interactive Enhancement Network,PEIE-Net),以逐步增强的方式减少图像增强过程中特征细节的丢失.具体来说,该网络由图像增强分支与边缘预测分支组成.在图像增强分支的每个增强阶段中,设计自注意力调制预测模块提取全局信息,用于对通道调制模块和多尺度复原模块进行自适应调制.在边缘预测分支中,设计空频域特征变换模块提取边缘感知信息,既用于预测高质量图像的边缘,又与图像增强分支的特征进行融合,以此模拟人类视觉系统在不同感知之间的交互.此外,本文还提出了场景亮度估计损失对多个渐进式增强阶段进行协调.在合成与真实数据集上的实验验证了本文方法在增强暗光、噪声、模糊退化图像方面的有效性与先进性,并可用于暗光图像增强与超分辨率任务.
Images captured in low-light scenes are susceptible to multiple degradations such as darkness,noise,and blur,resulting in poor visibility and visual perception.Multi-degraded low-light image enhancement poses challenges to existing image enhancement methods as follows:on the one hand,low-light image enhancement or deblurring methods cannot handle all three types of degradation,and the effect of the combination strategy is limited by the increased computational cost and error accumulation.On the other hand,the existing multi-degraded low-light image enhancement method adopts the strategy of enhancing brightness first and then removing blur,and this sequential processing manner increases the risk of losing feature cues and is not conducive to detail recovery.To cope with the above challenges,this paper proposes the progressive edge-aware interactive enhancement network(PEIE-Net),which reduces the loss of feature details by designing a step-by-step enhancement process.Specifically,our network consists of an image enhancement branch and an edge information prediction branch.In each enhancement stage of the image enhancement branch,a self-attention modulation prediction module is designed to extract the global information,which is used for adaptive modulation in the channel modulation module and multi-scale restoration module.In the edge information prediction branch,the spatial-frequency domain feature transformation module is developed to extract the edge perceptual information.The edge perceptual information is used to predict the edges of high-quality images while also fused with the image enhancement features,simulating the interaction between different perceptions within the human visual system.In addition,we propose scene brightness estimation loss to coordinate the multiple progressive enhancement stages.Experiments on synthetic and real datasets demonstrate the effectiveness and sophistication of our method for enhancing low-light,noisy,and blur-degraded images,and can be used for low-light image enhancement and super-resolution tasks.
作者
李悦洲
牛玉贞
李富晟
柯逍
施逸青
LI Yue-zhou;NIU Yu-zhen;LI Fu-sheng;KE Xiao;SHI Yi-qing(College of Computer and Data Science,Fuzhou University,Fuzhou,Fujian 350108,China;College of Photonic and Electronic Engineering,Fujian Normal University,Fuzhou,Fujian 350117,China)
出处
《电子学报》
北大核心
2025年第3期926-940,共15页
Acta Electronica Sinica
基金
国家自然科学基金(No.U21A20472,No.62072110,No.61972097)
福建省科技重大专项(No.2021HZ022007)
福建省自然科学基金(No.2023J01067,No.2020J01494)
福建省科技厅高校产学合作项目(No.2021H6022)。
关键词
图像增强
多退化图像
暗光增强
去模糊
特征调制
image enhancement
multi-degraded image
low-light enhancement
deblurring
feature modulation
