摘要
针对目前低照度图像增强方法,由于缺乏足够的低照度数据集,容易出现模型泛化能力差和过拟合的问题,本文提出一种改进的生成对抗网络低照度图像增强方法.在保持原生成对抗网络主干不变的基础上做了系列改进:首先在真实图像与生成图像中应用可微增强对数据进行扩容.其次,为了更有效地提取特征,提升训练效果,生成网络添加了残差结构和scSE注意力机制.此外,为了提高模型训练的稳定性,网络优化过程使用huber函数平滑损失值.实验表明,本文所提方法与主流低照度增强方法相比,在多项评价指标方面具有明显优势.
For the current low illumination image enhancement methods,due to the lack of sufficient low illumination datasets,it is easy to have the problems of poor model generalization ability and over fitting.This paper proposes an improved low illumination image enhancement method for Generative Adversarial Nets.A series of improvements have been made on the basis of maintaining the original backbone Generative Adversarial Nets.Firstly,the real image and the generated image are introduced with differentiable augmentation to expand the data capacity.Secondly,in order to extract more features and improve the training effect,residual structure and scSE attention mechanism are added to the generation network.In addition,for the stability of model training,huber function is used to smooth the loss value in the generation network optimization process.Experiments show that compared with the mainstream low illumination enhancement algorithm,it has obvious advantages in SSIM,PSNR,L1 Loss evaluation indicators.
作者
陈爱国
张翔宇
邹明杰
蒋亦樟
CHEN Aiguo;ZHANG Xiangyu;ZOU Mingjie;JIANG Yizhang(School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi 214000,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2024年第5期1109-1115,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(62171203)资助。
作者简介
陈爱国,男,1975年生,博士,副教授,CCF会员,研究方向为计算机视觉等,E-mail:529754824@qq.com;张翔宇,男,1998年生,硕士研究生,研究方向为图像处理、生成对抗网络;邹明杰,男,1998年生,硕士研究生,研究方向为计算机视觉;蒋亦樟,男,1988年生,博士,副教授,CCF高级会员,研究方向为模式识别、计算智能及其智慧医疗方面应用.
