谢林芳,张旭东,孙 锐,范之国,党天一.基于多对一映射生成对抗网络的颜色恒常性算法[J].电子测量与仪器学报,2022,36(4):124-135 |
基于多对一映射生成对抗网络的颜色恒常性算法 |
MTO-GAN: Learning many-to-one mappings for color constancy |
|
DOI: |
中文关键词: 颜色恒常性 单光源 生成对抗网络 图像映射 |
英文关键词:color constancy single illumination GAN image mapping |
基金项目:国家自然科学基金(61876057,61971177)、安徽省重点研发计划 科技强警专项(202004d07020012)项目资助 |
|
|
摘要点击次数: 513 |
全文下载次数: 734 |
中文摘要: |
颜色恒常性是计算机视觉的重要研究方向,但在目前的研究中,大多数算法专注于单光源均匀分布的情况,光源非均匀
分布的问题一直没有得到很好的解决。 针对该问题,在单光源非均匀分布的情况下,将颜色恒常性转换成一个多对一映射任
务,提出了一种基于生成对抗网络直接校正的方法。 该方法根据颜色恒常性的特性,将图像拆解成内容编码和光源编码,改变
光源编码为目标光源编码,重组生成目标光源下的图像。 为了使非标准光源更加多样化,加入光源采样模块,帮助网络学习到
更为丰富的光源信息,实现多对一映射。 同时,为了在输入不同的光源编码时可以引导图像映射到不同的光源下,加入光源监
督模块以区分具有不同光源的图像,帮助光源转换模块更好地将内容编码和特定光源编码结合起来,生成目标图像,实现颜色
恒常性。 同时,针对本文任务,在现存的数据集上渲染非均匀分布的光源,构造了单光源非均匀分布的数据集。 实验结果表明,
本文方法较好地解决了光源非均匀分布的问题,在非均匀数据集上超越了其他算法,最终生成的图像也更加接近标准光源下
图像。 |
英文摘要: |
Color constancy is an important research direction in computer vision, but most algorithms focus on uniform distribution of
single illuminant, and the problem of non-uniform distribution of illuminant has not been well solved. In order to solve this problem, a
direct correction method based on generative adversarial network is proposed to transform color constancy into a many-to-one mapping
task under the condition of non-uniform distribution of single illuminant. According to the characteristic of color constancy, the image is
divided into content code and illuminant code, and the image under the target illuminant is reconstructed by changing the illuminant code
to target illuminant code. At the same time, in order to make non-standard illuminant more diversified, the illuminant sampling module
is added to help the network learn more abundant illuminant information and realize many-to-one mapping. In order to guide images to be
mapped to different illuminants when different illuminant codes are input, the illuminant supervision module is added to distinguish
images with different illuminants, so as to help the illuminant conversion module better combine content coding with specific illuminant
coding to generate target images and achieve color constancy. At the same time, aiming at the task of this paper, the non-uniformly
distributed illuminant is rendered on the existing dataset, and the dataset with non-uniformly distributed single illuminant is constructed.
The experimental results show that the proposed method solves the problem of non-uniform distribution of illuminant well, surpasses other
algorithms in non-uniform dataset, and the final image is closer to the image under standard illuminant. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|