Hi Francis,
The Wolfram NN repository page for ImageIdentify
says
over 3 million training images and over 4,000 classes of objects (not
publicly available).
The network architecture is available on the NN repository page, you can compare it to VGG19 to see what is different. You could try running the ImageIdentify NN against the ImageNet images and compare performance with VGG19.
You could also try contacting Wolfram Support and asking for more information on the training set.