> 文章列表 > Learning Tone Curves for Local Image Enhancement

Learning Tone Curves for Local Image Enhancement

Learning Tone Curves for Local Image Enhancement

作者 LUXI ZHAO , ABDELRAHMAN ABDELHAMED , AND MICHAEL S. BROWN

论文比较清晰易懂。

就是图像分8∗8=648*8=6488=64个patch, 卷积网络为RGB三个通道预测 3∗643*64364 个 look up table,
就是每个patch 3个1D lut.
Learning Tone Curves for Local Image Enhancement

然后每个patch的中心直接用1D lut, 其他部分像素用周边lut插值得到,避免引入aritfact.
Learning Tone Curves for Local Image Enhancement

该论文研究的内容在ISP中的位置:
Learning Tone Curves for Local Image Enhancement
损失函数是 L2 + 浅层vgg特征损失
Learning Tone Curves for Local Image Enhancement
code and blog:

https://github.com/SamsungLabs/ltmnet
https://zhuanlan.zhihu.com/p/558296637