Lecture 12 | Computational Photography 1¶

High Dynamic Range Imaging (HDR)¶

我们希望亮的和暗的地方细节都很丰富

曝光 exposure = Gain增益 * Irradiance辐照度 * Time曝光时间

分别取决于……

• 快门
• 光圈
• ISO 感光度
  • 高：更亮、更多噪声

Dynamic range

The ratio between the largest and smallest values of a certain quantity

HRD: 不同亮度都拍一些，然后合成

12 to 8 -> tone mapping

Gamma compression

Deblurring¶

模糊的两个主要原因：defocus and motion blur

blurred image = clear image * blur kernel

deblurring = deconvolution

• Non-blind image deconvolution, NBID
• Blind image deconvolution, BID

inverse filter

wiener filter

Colorization¶

最关键的问题是，如何决定上的颜色

主要有两类方法：

1. sample-based colorization
2. interactive colorization

Sample-based¶

...

Interactive¶

...

对于视频

Colorful Image Colorization

GAN (Generative Adversarial Network)

• 两个一起训练，min-max problem
• GAN 很难收敛，不好调
• adversarial loss

more Image synthesis tasks¶

• image to image translation
  • pix2pix gan
• style transfer
• text-to-photo
• image dehazing
• customized gaming

GAN 的改进：Cycle-GAN，解决了缺少成对的训练数据的问题

Super Resolution¶

up sampling

bi-cubic

super resolution using GAN

激励

评论