(Day 30) KCSE 2024 day 1 and Face recognition & Neural Style transfer

Ivan Ivanov · January 31, 2024

deep-learning cnn reading-research

Hello :) Today is Day 30!

A quick summary of today:

covered the last bit of the CNN course
first day of KCSE 2024

Firstly, the last bit of the CNN course

It was about face recognition and neural style transfers.

Face recognition

You need to be aware of the difference between recognition and verification. Verification is the process of showing a face and verifying that it is the person by comparing it with the image of the person in the system. Recognition is the process of showing a face and verifying that a person’s face is among all the pictures in the system.

How do we evaluate a face recognition model?

How do we compute similarity?

You can give all the pictures to the same model, compare the data of the last Dense layer, and calculate the loss function in that way. There is also a variable alpha. This is to use three pictures at a time and anchor A (self image), positive P (other self image), and negative N (fake image). The accuracy of the model depends on the size of max between 0 and the difference between the absolute values of A-P, and A-N, and adding alpha. High means that the model is more certain when the model recognizes or rejects the face.

We can also use logistic regression

Neural style transfer

We give images C and S to a model, and the output is a image C with the style of S applied to it. The loss function is to calculate the difference between C and S image features.