The Numediart Institute has a long expertise in the analysis of human visual perception and its attention which allows to filter the acquired information. Attention has a generic component which is not based on any specific learning and a component which depends on intentions, knowledge or emotions of a specific person. With the arrival of deep learning models (DNN) high-level information suddenly became available. This project focuses on how to integrate those high-level information (memorability, aesthetics, object, concepts or action detection) with the low-level reflex information in a way which is coherent with the human visual system. This project is part of a larger domain of artificial intelligence and will focus on deep learning.