摘要
针对低照度图像存在识别度不高、亮度低、信噪比低和细节模糊等问题,提出了一种非下采样剪切波变换(NSST)域结合生成对抗网络(GAN)的低照度图像增强方法。首先,收集弱光图像和正常光图像数据集,将图像进行RGB颜色空间到HSV颜色空间的变换处理,保持色度、饱和度分量不变,对亮度分量进行NSST多尺度分解,利用分解得到的低通子带图像构建训练集;其次,构建基于GAN的低频子带图像增强模型,并利用低频子带图像训练集对模型进行训练;然后,对待处理的低照度图像进行NSST分解,利用训练的模型增强低频子带图像,利用尺度相关系数去除各高频方向子带噪声,并通过非线性增益函数增强边缘系数;最后,将增强处理后的低频、高频子带图像进行NSST重构,并将重构图像恢复至RGB颜色空间。所提方法与常见的方法相比,就低照度图像增强而言,结构相似度平均提升了3.89%,均方误差平均降低了1.03%,且在对噪声图像增强时,峰值信噪比和连续边缘像素比保持在21 dB和88%以上。实验结果表明,所提方法不论从视觉效果还是图像质量客观评价指标上较常见方法都有较大提升,能有效改善低照度图像的低质问题,为后续的图像处理分析奠定基础。
Low illumination image has a number of issues,such as low recognition,low brightness,low resolution,low signal-to-noise ratio and blurred details.Therefore,a low-light image enhancement method combined with generative adversarial networks(GAN)in nonsubsampled shearlet transform(NSST)domain is proposed.First,low-light image and normal light image datasets are collected,the images are processed by RGB to HSV spatial transformation,the Hue and the Saturation components are unchanged,the Value components are decomposed at multiple scales by NSST,and the decomposed low-pass subband images are used to construct training set.Second,a low-frequency subband image enhancement model based on GAN is constructed,and the low-frequency subband image training set is used to train the model.Then,the low-illumination image to be processed is decomposed by NSST,the trained model is used to enhance the low-frequency subband image,the scale correlation coefficient is used to remove noise for each high-frequency direction subband,and the edge coefficient is enhanced by the nonlinear gain function.Finally,NSST reconstruction is performed on the low-frequency and high-frequency subband images after enhanced processing,and the reconstructed images are restored to RGB space.In terms of low-light image enhancement,compared to common methods,the results obtained by the proposed method show an average improvement of 3.89%in structural similarity and an average reduction of 1.03%in mean squared error,and when the noisy images are enhanced,the peak signal to noise ratio and continuous edge pixel ratio remain above 21 dB and 88%,respectively.The experimental results show that both visual effect and objective evaluation index of image quality of the proposed method are greatly improved compared to the common methods,which can effectively improve the low-quality problem of low-light images,and lay the foundation for the subsequent image processing analysis.
作者
施雯玲
廖一鹏
许志猛
严欣
朱坤华
Shi Wenling;Liao Yipeng;Xu Zhimeng;Yan Xin;Zhu Kunhua(College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,Fujian,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第24期120-130,共11页
Laser & Optoelectronics Progress
基金
国家自然科学基金(62271149,62271151)
福建省自然科学基金(2019J01224)。
关键词
低照度图像增强
非下采样剪切波变换
生成对抗网络
图像去噪
图像边缘增强
low-light image enhancement
nonsubsampled shearlet transform
generative adversarial network
image denoising
image edge enhancement
