Preferably a trained model that can be used directly, like the Google API:https://cloud.google.com/vision/docs/labels