The idea is to use the VGG16 network pretrained on a large dataset such as ImageNet. Note that training could be rather computationally expensive and therefore makes sense to reuse an already pretrained network:

So, how to use a VGG16? Keras makes it easy as the library has a standard VGG16 application available as a library and precomputed weights are automatically downloaded. Note that we explicitly omit the last layer and replace it with our custom layer, which will be fine-tuned on the top of the prebuilt VGG16. In this example, you will learn how to classify dogs and cats images available from Kaggle.