What kind of features are extracted with the AlexNet layers?

Question is regarding the method where the fully connected layer 'fc7' from the pre-trained AlexNet is used to classify images, method can be seen here https://se.mathworks.com/help/nnet/ref/alexnet.html.

What kind of features is it actually extracting?

I used this method on images of paintings on two different painters. First using about 150 training images from each artist and then determining the correct artist on 40 test images. This works really well, about 85-90% correct classification and I would like to know why it works so well. The features are stored in a 300x4096 matrix.