文章目录
Representative Color Transform for Image Enhancement
作者:Hanul Kim1, Su-Min Choi2, Chang-Su Kim3, Yeong Jun Koh
单位:Seoul National University of Science and Technology 2Chungnam National University 3Korea University
Abstract
前人方法都是encode-decode方式,丢失细节;密集转化也限制颜色空间的迁移效果;
本文使用颜色迁移表征(RCT)表征颜色变化,根据输入和表征颜色相似性增强颜色,得到更好效果;
RCT determines different representative colors specialized in in- put images and estimates transformed colors for the repre- sentative colors. It then determines enhanced colors us- ing these transformed colors based on the similarity be- tween input and representative colors.
Introduction
- 问题
First, details of the input im- age are not preserved in the up-sampling process of the de- coder, even though they employ skip-connections. Second, these approaches train networks with fixed input size, which makes it difficult to enhance images of arbitrary spatial res- olutions in the inference phase.
1.使用上采样无法保证解码器还原细节特征;
2.固定尺寸输入让输入图片有限制;
还有一种方法:全局参数估计,不需要上采样,使用RGB, CIELab,LUTS等方式,但不同通道之间无法分开训练颜色迁移;
- 本文使用方法
RCT学习大规模色彩迁移,encode+特征表征=大规模色彩迁移能力
Related work
- 数据集MIT-Adobe 5K,深度学习开始的方法直接学习像素级,端到端,但效果不行;
- 逐渐出现encoder-decoder方法,从GAN到预测光流网络,到频率分解方法;
- 全局参数估计:使用密集迁移方程、通道密集迁移、强化学习、3D LUTS,预定义无法有效迁移;
Method
得到模型结构,代码如下
实验
在四个实验数据集上测试,达到不错结果;
Conclusion
使用表征和颜色转换特征,利用全局和局部特征融合,得到对应颜色矩阵,提高色彩强化效果。