IMFs和改进GAN的两幅大曝光率比图像高动态融合
Two large-exposure-ratio images high dynamic fusion based on IMFs and improved GAN
吴宇 1方红萍 1伍世虔1
作者信息
- 1. 武汉科技大学信息科学与工程学院,湖北武汉 430081;武汉科技大学机器人与智能系统研究院,湖北武汉 430081
- 折叠
摘要
针对两幅大曝光率比图像高动态融合时颜色和明暗对比度失真的问题,提出一种IMFs和改进GAN的高动态融合算法.利用强度映射函数(IMFs)插值一张中间虚拟曝光图像;引入曝光感知补偿模块EACB提取可靠区域特征,设计改进GAN建模图像残差,定义渐进学习策略保证GAN稳定收敛,实现中间虚拟曝光图像增强;基于3张图像实现多尺度曝光融合.实验结果表明,针对曝光间隔4EV的高低曝光图像集,算法能有效抑制颜色和明暗度对比失真,保留纹理细节,客观指标MEF-SSIM优于经典MEF算法.
Abstract
According to color and shading contrast distortion in the high dynamic fusion for two large exposure ratio images,a high dynamic fusion algorithm combining IMFs and improved GAN was proposed.An initial intermediate virtual exposure image was obtained by intensity mapping functions(IMFs).An improved GAN driven by IMFs,which accelerated the convergence and enhanced the details of the intermediate virtual exposure image,was designed to model the residual image.EACB module was introduced to guide the network to extract features from the well-exposed reliable area.A proposed generation network progres-sive learning strategy ensured fast and effective convergence of the GAN to obtain high-quality intermediate virtual exposure ima-ges.A classical exposure fusion algorithm was used to achieve high dynamic fusion based on above three images.Experimental results on high and low exposure image dataset with 4EV exposure interval show that the relative brightness and more details are preserved much better in the fusion result images.The MEF-SSIM of the proposed algorithm surpasses that of existing classical MEF algorithms.
关键词
两个大曝光比图像/高动态融合/曝光插值/强度映射函数/改进生成对抗网络/曝光感知补偿块/多尺度曝光融合Key words
two large-exposure-ratio images/high dynamic fusion/exposure interpolation/intensity mapping functions/im-proved GAN/exposedness aware compensation block/multiscale exposure fusion引用本文复制引用
基金项目
湖北省自然科学基金青年基金项目(2022CFB676)
出版年
2024
NETL
NSTL
万方数据