The idea of the similarity function is to compare what the network "sees" in our image gluing to what it "sees" in the target picture. To do this, the idea is to take a publicly available pretrained neural network and compare the output of some of the convolution features.
This formula is used to quantify how similar are the contents of two images, y the one we are trying to reproduce, and ŷ the one we are building. Φ(y) is a vector that represents the content of y.
I will build a notebook that demonstrates this concept.
The idea of the similarity function is to compare what the network "sees" in our image gluing to what it "sees" in the target picture. To do this, the idea is to take a publicly available pretrained neural network and compare the output of some of the convolution features.
The idea is presented in the equation (2) of the following paper Perceptual Losses for Real-Time Style Transfer and Super-Resolution:
This formula is used to quantify how similar are the contents of two images,
y
the one we are trying to reproduce, and ŷ the one we are building.Φ(y)
is a vector that represents the content ofy
.I will build a notebook that demonstrates this concept.