Fundamentals of face recognition (facenet)

Main ideas for reference ：

• FaceNet The paper :FaceNet: A Unified Embedding for Face Recognition and Clustering
• Wu Enda video
• Keras bosses

Face verification VS Face recoginition

For face verification , We can regard it as a binary classification problem , But face recognition has developed to the present , We still use face recognition to solve .

Encoder

• We enter a picture into encoder, Through a series of Networks , Output a unique password , This string of passwords compiles this person .
• When we compare , Pass the current picture through encoder Code to get a string of passwords , Compare this string of passwords with all the passwords in the database . If $f(key,ke{y}_{i_{database}})\le \alpha$, We think this person is the person in the database .

Triplet Loss

That's the question ： How do we train this network （Encoder）？

We divide the data into three groups .
$$\begin{gathered} （Anchor,Positive,Negative） \\ AnchorsurfaceinIPeopleNeed\; to\; bewantknowledgeotherOfthisindividualpeople,PositivesurfaceinThepeopleOfanotherOneZhangchartslice \\ ,NegativesurfaceinanotherOneindividualpeopleOfchartslice. \end{gathered}$$
$$\begin{gathered} IPeoplesetThe\; righteousTripletLoss： \\ Max(||(ke{y}_{Anchor})-(ke{y}_{Positive})||-||(ke{y}_{Anchor}) \\ -(ke{y}_{Negative})||+\alpha ,0) \end{gathered}$$
We are optimizing Encoder When , We need different photos of the same person Encoder The compiled key The gap should be as small as possible , At the same time, I hope that the pictures of different people will be compiled key The gap should be as big as possible .

As for the neural network in the middle , You can write your own , You can even use the previous image classification VGG perhaps ConvNet.