week04.Ng's CNN Course

CNN 2017.12

Face Recognition

1. What is face recognition?

• Baidu's face recognition demo

• face verification - $1:1$

• face recognition - $1:K$

2. One-Shot learning

• one-shot learning

• Learning a "similarity" function

  $d(img1, img2)$ = degree of difference between images
  if $d(img1, img2) \le \tau$
  if $d(img1, img2) > \tau$

3. Face recognition

[Taigman et. al., 2014. DeepFace closing the gap to human level performance]

• siamese network architecture

  $x^{(1)} \rightarrow ConvNets \rightarrow f(x^{(1)})$
  encoding of the image
  $x^{(2)} \rightarrow ConvNets \rightarrow f(x^{(2)})$

4. Triplet Loss

[2015,FaceNet:A unified embedding for face recognition and clustering]

• Learning Objective

  anchor image $A$

  positive $P$

  negative $N$

  want : $d(A, P) \le d(A, N)$
  $d(A, P) \le d(A, N) -\alpha<0$
  $\alpha$ is called margin
  确保总是等于0的情况不出现，并且使得负样本和正样本分别距anchor之间距离的差存在一定的margin。

• Loss function

  给定三张图片$A,P,N$
  

  $$\begin{align*} L(A,P,N) = \max( \parallel f(A) - f(P) \parallel^2 - \parallel f(A) - f(N) \parallel ^ 2 + \alpha , 0 ) \end{align*}$$

  有多张图片的
  

  $$\begin{align*} J = \sum_{i=1}^m L(A^{(i)}, P^{(i)}, N^{(i)}) \end{align*}$$

  training set: 10k pictures of 1k persons

• Choosing the tripets $A,P,N$

  

  选择相近的P、N来训练以提升网络的效能

5. Face verification and Binary Classifiction

• Learning the similarity function

  将两张脸同时输入网络中(e.g. siamese network), 得到两个输出，然后同时输入一个logistic regression unit，得到最后的结果。全部相同为1，不同为0. 转化为一个2分类问题。

  • $\chi^2$ dissimilarity

Neural style transfer

1. what is neural style transfer?

2. What are deep ConvNets Learning？

如何进行可视化？

在一个隐藏层中，选取使它的activation最大的几个图像的patch，看看是什么。

• visualizing deep layers

3. Cost function？

• Neural style transfer cost function？

  

  给定一张content image C
  再给定一张style image S
  得到generated image G
  定义cost function $J$

• Find the generated image G

  先随机初始化$G$为一定大小
  利用梯度下降法不断迭代$G$

4. Content Cost Function

5. Style Cost Function

• Gram Matrix

  利用Gram矩阵来衡量风格的相似度。

5. 1D and 3D Generalization

• 之前讨论的都是2D的图像

  1D的卷积

  3D的数据，例如X-ray照片，有多个切片

  卷积在不同信号上的应用