We present a deep learning method for supervised image classification, integrating knowledge formalized as a graph. We introduce a cost function combining the classical cross-entropy used in deep learning and an original function based on the representation of nodes after an embedding of the knowledge graph. The knowledge is only used during the learning phase and is not necessary to evaluate an example. Experiments on several image databases show an improvement of the performance compared to state-of-the-art: in comparison with classical deep learning algorithms, and with recent algorithms also integrating knowledge represented by a graph.