摘要
深度Retinex-Net算法利用了低/正常光图像具有相同反射率的约束条件,以数据驱动的方式实现了弱光照图像的增强。该算法解决了传统图像增强算法非线性表达能力不强以及增强后的图像不自然等一系列问题。但在该算法中分解出的照度分量模糊且不够平滑,以及对反射分量处理时采用的BM3D去噪操作没有考虑噪声对不同光照区域的影响,导致图像增强效果一般。鉴于Retinex-Net算法的局限性,提出了一种基于Retinex模型的弱光照图像增强算法。为了更准确地计算分解出照度分量的估计值,提出了一个照度分量平滑度损失函数来更好地学习分解的过程,并使用U-Net网络结构对反射分量中存在的噪声进行去噪,最后将两者进行融合得到增强后的图像。实验结果表明,该算法不仅能有效地提高主观视觉效果上的图像对比度、亮度和色彩饱和度,在客观评价指标上如PSNR和SSIM也均得到了进一步提高。
The deep Retinex-Net algorithm takes advantage of the constraints that low/normal light images have the same reflectivity and light smoothness and realizes the enhancement of weak light images in a data-driven manner.This algorithm solves a series of problems of traditional image enhancement algorithms such as the non-linear expression ability and the unnatural feature of enhanced images.However,the illuminance component decomposed in this algorithm is fuzzy and not smooth enough,and the BM3D denoising operation used in the processing of the reflection component does not consider the effect of noise on different illumination areas,resulting in a general image enhancement effect.In view of the limitations of the Retinex-Net algorithm,a weakly illuminated image enhancement algorithm based on the Retinex model is proposed.In order to calculate the estimated value of the decomposed illuminance component more accurately,a loss function of the illuminance component smoothness is proposed to better learn the decomposition process,and the U-Net network structure is used to denoise in the reflected component.These two measures are fused to obtain the enhanced image.The experiment shows that the proposed algorithm can effectively improve the subjective quality such as the contrast,brightness and color saturation of enhanced images and the objective quality such as PSNR and SSIM.
作者
李博文
唐贵进
崔子冠
LI Bo-wen;TANG Gui-jin;CUI Zi-guan(Jiangsu Key Laboratory of Image Processing and Image Communication,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
出处
《计算机技术与发展》
2021年第5期79-84,共6页
Computer Technology and Development
基金
国家自然科学基金(61501260)
江苏省科协提升计划项目(TJ215039)
南京邮电大学科研基金项目(NY219076)。
