[machine learning notes] [face recognition] deeplearning ai course4 4th week programming

 The goal is ：
	1.  Use pre trained models , Conduct 
	1.1  Face verification 
	1.2  Face recognition 

Reference from ：
【 chinese 】【 Wu Enda's after-school programming homework 】Course 4 - Convolutional neural networks - Work in the fourth week - Face recognition and neural style transformation

Face verification

Give a group 【 Face and identity 】, Look up... In the database ,【 Face 】 and 【 identity 】 match .

Send two pictures into CNN Online code become 128 Dimension vector .
Compare the two code Between distance , The smaller the distance , The more similar

CNN The Internet

Use FaceNet Model （ Others have been trained ）

CNN Loss function of

Triplet loss

• A yes “Anchor”, It's an image of a person .
• P yes “Positive”, Is relative to “Anchor” Another image of the same person .
• N yes “Negative”, Is relative to “Anchor” Another image of a different person .

A-P The square of , Add each dimension + $\alpha$ < A-N The square of , Add each dimension
$$\sum _i^n||f(A_i)-f(P_i)|{|}^2+\alpha <\sum _i^n||f(A_i)-f(N_i)|{|}^2$$
A And P Between distance Be as small as possible ,A And N Between distance Try to be as big as possible .
among ,$\alpha$ yes 【 identical 】 and 【 Different 】 At least the distinction between images .

---

Transposition to ,
$$\sum _i^n||f(A_i)-f(P_i)|{|}^2-\sum _i^n||f(A_i)-f(N_i)|{|}^2+\alpha <0$$
We define , Loss function （ Triplet loss ）:
$$J(A,P,N)=max(\sum _i^n||f(A_i)-f(P_i)|{|}^2-\sum _i^n||f(A_i)-f(N_i)|{|}^2+\alpha ,0)$$
send J As small as possible .

Implementation verification

Input 【 Face 】 and 【 identity 】. We need to verify that , The 【 Face 】 Is it this 【 identity 】.

---

We can store all the data in the database in advance 【 Face 】 Encoding , And use Dictionaries take 【 Face 】 and 【 identity 】 One-to-one correspondence .

after , Whenever the camera takes a picture , We just need to put it in CNN Online code , And then This code And database Code of corresponding identity Calculate both distance . if distance dist Less than threshold The verification is successful .

Face recognition

Input 【 Face 】. We need to verify that , The 【 Face 】 Is it in the database .

---

The practice is similar Face verification , In advance, code Good database all 【 Face 】

Traverse the coding of all faces in the database , Calculate the distance between the two . Take the minimum value of all distances .

If it's time to Minimum distance Less than threshold , Then the verification is successful .