Visual attention with deep neural networks

Animals use attentional mechanisms for being able to process enormous amount of sensory input in real time. Analogously, computerised systems could take advantage of similar techniques for achieving better timing performance. Visual attentional control uses bottom-up and top-down saliency maps for establishing the most relevant locations to observe. This article presents a novel fully-learnt unbiassed biologically plausible algorithm for computing both feature based and proto-object saliency maps, using a deep convolutional neural network simply trained on a single-class classification task, by unveiling its internal attentional apparatus. We are able to process 2 megapixels (MPs) colour images in real-time, i.e. at more than 10 frames per second, producing a 2MP map of interest.